fortran user's guide (primehpc fx100)fumie/internal/riken... · 2017. 4. 6. · preface...

J2UL-1865-02ENZ0(00)November 2015

FUJITSU SoftwareTechnical Computing Suite V2.0

Fortran User's Guide(PRIMEHPC FX100)

Preface

Purpose of This Manual

This manual explains how to use the Fortran system (called "this system" in this manual) that runs on the Linux operating system. It alsodescribes the following aspects of the system:

- Fortran programming

- Restrictions

Intended Readers

This manual was written for people who write Fortran programs and process them using this system.

This manual assumes that readers know how to use Linux commands, file manipulation and shell programming.

This manual assumes an understanding of the Fortran language specification. For details of the Fortran language specification, refer to"Fortran Language Reference" on online manual, the Fortran language standard, or commercial books.

Organization of This Manual

This manual is organized as follows:

Chapter 1 Overview of the Fortran System

This chapter provides an overview of how to use this system to process Fortran programs.

Chapter 2 Compiling and Linking Fortran Programs

This chapter describes how to compile and link Fortran programs.

Chapter 3 Executing Fortran Programs

This chapter describes how to execute Fortran programs.

Chapter 4 Messages and Listings

This chapter describes the system messages and listings produced as a result of compiling a Fortran source program or executing aFortran program.

Chapter 5 Data Types, Kinds, and Representations

This chapter describes the Fortran data types that can be used in this system.

Chapter 6 Notes on Programming

This chapter describes some of notes on programming on this system. Input/output features are described in "Chapter 7 Input/outputProcessing".

Chapter 7 Input/output Processing

This chapter describes files processed by Input/output statements and the necessary basic properties of Fortran records and files forusing input/output statements.

Chapter 8 Debugging

This chapter describes the debugging functions provided by this system and how to operate them.

Chapter 9 Optimization Functions

This chapter describes functions to use optimization functions effectively, as well as some points to note.

Chapter 10 Fortran Modules

This chapter describes some of notes on compiling programs including modules.

Chapter 11 Mixed Language Programming

This chapter describes the specification and notes when linking a Fortran program and a C/C++ program.

- i -

Chapter 12 Multiprocessing

This chapter describes how parallel processing is performed for a Fortran program in this system.

Appendix A Limits and Restrictions

This appendix describes limits and restrictions on programing using this system.

Appendix B Printing Control Command

This appendix describes printing control commands provided by this system.

Appendix C Endian Convert Command

This appendix describes endian convert commands provided by this system.

Appendix D Job Operation Software

This appendix describes information on the compilation and execution processes for using the high-speed facility on the Job OperationSoftware.

Appendix E Preprocessor

This appendix describes the C language preprocessor and the Fortran preprocessor.

Appendix F Notes on Migration from FX10 System to FX100 System

This appendix describes notes on migrating from FX10 system to FX100 system.

Appendix G Compatibility Information (FX10 System)

This appendix describes compatibility information as notes on migrating.

Appendix H Compatibility Information (FX100 System)

This appendix describes compatibility information as notes on migrating.

Related Manuals

The following manuals are related to this manual.

- Fortran Language Reference

- Fortran Compiler Messages

- Fortran/C/C++ Runtime Messages

- Fortran User's Guide Additional Volume COARRAY

Notes of This Manual

The code examples of optimization in this manual are conceptual source that complements each explanation of functions. When the codeexamples are compiled and executed, optimizations may not work as expected. This is because optimizations depend on compilationoptions and other conditions.

Notation Used in This Manual

Syntax Description Symbols

A syntax description symbol is a symbol that has a specific meaning when used to describe syntax. The following symbols are usedin this manual:

Symbol name Symbol Explanation

Selection symbols

{ }Only one of the items enclosed in the braces must be selected (itemsare listed vertically).

|Multiple items are enumerated by this delimiter (items are listedhorizontally).

Option symbol [ ]An item enclosed in brackets can be omitted. This symbol includes themeaning of the selection symbol "{}".

- ii -

Symbol name Symbol Explanation

Repeat symbol ...The item immediately preceding the ellipsis can be specified repeatedlyin the syntax.

Parallel

The way of parallelization is thread parallelization unless otherwise described.

Export Controls

Exportation/release of this document may require necessary procedures in accordance with the regulations of your resident countryand/or US export control laws.

Trademarks

- Oracle and Java are registered trademarks of Oracle and/or its affiliates.

- All SPARC trademarks are used under license from SPARC International, Inc. and are trademarks or registered trademarks of thatcompany in the United States and other countries.

- OpenMP is a trademark of OpenMP Architecture Review Board.

- Linux is a trademark or registered trademark of Linus Torvalds in the United States and other countries.

- All other trademarks and product names are the property of their respective owners. The trademark notice symbol (TM, (R)) is notnecessarily added in the system name and the product name, etc. published in this material.

Date of Publication and Version

Version Manual code

November 2015, 2nd Version J2UL-1865-02ENZ0(00)

February 2015, Version 1.1 J2UL-1865-01ENZ0(01)

October 2014, 1st Version J2UL-1865-01ENZ0(00)

Copyright

Copyright FUJITSU LIMITED 2014-2015

Update History

Changes Location Version

Added the article "Parallel" to "Notation Used in This Manual". Preface 2nd Version

Added the article on the following option:

- -K{ openmp_tls | openmp_notls }

2.212.3.1.1.1

Added articles on the following options:

- -K{ FLTLD | NOFLTLD }

- -K{ HPC_ACE | HPC_ACE2 }

- -K{ intentopt | nointentopt }

- -K{ lto | nolto }

- -K{ simd_separate_stride | simd_noseparate_stride }

- -N{ coarray | nocoarray }

2.2

Changed articles on the following options:

- iii -


- -g

- -K{ fed | nofed }

- -Kfp_relaxed

- -K{ nf | nonf }

- -Knoalias[=spec]

- -Ksimd=auto

- -NRtrap

- -Nsetvalue=struct

- -K{ uxsimd | nouxsimd }

- -S

- -W

Changed the article of the -Lu option. 3.3

Added the article on the environment variable FLIB_USE_STOPCODE. 3.8

Added the article on "Notes for Execution" 3.9.1

Added articles of the following notes:

- Link Time Optimization

- #line directives

4.1.2.3

Changed articles in response to the specification addition of "ERROR STOP statement". 4.24.2.16.5.78.1.8

Changed the article on the STATUS= specifier. 7.5.1

Changed the article. 8.2.1.3

Corrected the code example of List Vector Conversion. 9.1.1.7.3

Added the article "SIMD with Redundant Executions for the SIMD Width". 9.1.1.7.4

Changed the article. 9.1.2.6

Added the article "Link Time Optimization". 9.1.2.10

Changed the article. 9.7.2

Added articles on the following optimization control specifiers:

- { FLTLD | NOFLTLD }

- SIMD(ALIGNED)

- SIMD(UNALIGNED)

- { SIMD_REDUNDANT_VL(n) | SIMD_NOREDUNDANT_VL }

- UXSIMD(ALIGNED)

- UXSIMD(UNALIGNED)

9.10.4

Changed articles on the following optimization control specifiers:

- FP_RELAXED

- { NF | NONF }

- NOALIAS

- iv -


- SIMD

- UXSIMD

Corrected the code example of the optimization control specifier NOVREC.

Added the description in restrictions of the target variable of the dimension shift of the arraydeclaration.

9.14.2

Added the description in effects of merging array variables. 9.15.1

Added the description in restrictions of merging local array variables. 9.15.2

Added the article "Example of SIMD(ALIGNED) specifier". 9.18.1

Corrected the article "Notes when using service routines". 12.2.1.2.712.3.1.2.3

Corrected the article on the environment variable OMP_PROC_BIND. 12.3.2

Changed the article on BIND attribute. 12.3.3.7

Added the article "Derived type variable with length type parameter" to OpenMP specificationsand restrictions.

12.3.3.11

Added BLOCK construct and CRITICAL construct in limits of total nesting level. A.2

Updated notes on migration. Appendix F

Added that the specification changes of the -K{ FLTLD | NOFLTLD } options and theoptimization control specifiers { FLTLD | NOFLTLD } are only influenced when the -KHPC_ACE2 option is set.

F.7

Added that the specification changes of the optimization control specifiers SIMD and UXSIMDare only influenced when the -KHPC_ACE2 option is set.

F.12

Added the compatibility information. G.1H.1

Reviewed the code examples and command execution examples. -

Added the following articles:

- Related Manuals

- Notes of This Manual

Preface Version 1.1

Corrected the article of -Ay option. 2.2

Added articles on the following options:

- -K{ openmp_assume_norecurrence | openmp_noassume_norecurrence }

- -N{ cancel_overtime_compilation | nocancel_overtime_compilation }

Changed the option name from -K{ ordered_omp_reduction | noordered_omp_reduction } to -K{ openmp_ordered_reduction | openmp_noordered_reduction }.

2.212.3.1.1.1

Changed the article of -Lu option. 3.3

Changed the output of SIMD information. 4.1.2.14.1.2.2

Added the article on Fortran record which length is more than 2G bytes. 7.3.2.1

Deleted articles on the error number 1164 because the upper size limit of the output item list waschanged.

Chapter 8

Added the article of software pipelining. 9.1.1.6

Changed the article on the optimization control specifier ARRAY_MERGE. 9.10.4

Corrected the code example of the optimization control specifier NOVREC.

- v -


Added the article of loop distribution and automatic loop slicing. 12.2.3.1.6

Added the article on linkage with other multi-thread programs. 12.2.412.3.5

Changed the term from "hardware barrier in a cpu chip" to "on-chip hardware barrier". Appendix D

Updated notes on migration. Appendix F

Added the compatibility information. G.1

Added the compatibility information on FX100 system as the appendix. Appendix H

All rights reserved.The information in this manual is subject to change without notice.

- vi -

ContentsChapter 1 Overview of the Fortran System.............................................................................................................................. 1

1.1 Configuration of the Fortran System................................................................................................................................................... 11.2 How to use Fortran System..................................................................................................................................................................1

1.2.1 Preparation.................................................................................................................................................................................... 11.2.2 Compilation and Linking.............................................................................................................................................................. 11.2.3 File Names.................................................................................................................................................................................... 21.2.4 Some Compilation Options...........................................................................................................................................................21.2.5 Execution...................................................................................................................................................................................... 21.2.6 Debugging.....................................................................................................................................................................................31.2.7 Tuning...........................................................................................................................................................................................3

Chapter 2 Compiling and Linking Fortran Programs................................................................................................................ 42.1 Compile Command.............................................................................................................................................................................. 4

2.1.1 Compile Command Format.......................................................................................................................................................... 42.1.2 Compile Command Input File...................................................................................................................................................... 4

2.2 Compiler Options.................................................................................................................................................................................52.3 Compile Command Environment Variable....................................................................................................................................... 572.4 Compilation Profile File.................................................................................................................................................................... 582.5 Compiler Directive Lines...................................................................................................................................................................582.6 Return Values during Compilation.................................................................................................................................................... 59

Chapter 3 Executing Fortran Programs..................................................................................................................................603.1 Execution Command..........................................................................................................................................................................603.2 Execution Command Format............................................................................................................................................................. 603.3 Runtime Options................................................................................................................................................................................ 603.4 Execution Command Environment variable......................................................................................................................................653.5 Execution Profile File........................................................................................................................................................................ 653.6 Execution Command Return Values..................................................................................................................................................663.7 Standard Input, Output, and Error Output for Execution.................................................................................................................. 663.8 Using Environment Variables for Execution.....................................................................................................................................663.9 Notes for Execution........................................................................................................................................................................... 68

3.9.1 Notes for Execution of the Non-process Parallel Job on the FX100 System............................................................................. 68

Chapter 4 Messages and Listings.......................................................................................................................................... 704.1 Compilation Output........................................................................................................................................................................... 70

4.1.1 Compilation Diagnostic Messages..............................................................................................................................................704.1.2 Compilation Information............................................................................................................................................................ 71

4.1.2.1 Compilation Information Options........................................................................................................................................714.1.2.2 Example of Compilation Information..................................................................................................................................764.1.2.3 Notes on Compilation Information...................................................................................................................................... 85

4.2 Executable Program Output...............................................................................................................................................................864.2.1 Diagnostic Messages...................................................................................................................................................................874.2.2 Trace Back Map..........................................................................................................................................................................874.2.3 Error Summary........................................................................................................................................................................... 88

Chapter 5 Data Types, Kinds, and Representations.............................................................................................................. 895.1 Internal Data Representation............................................................................................................................................................. 89

5.1.1 Integer Type Data....................................................................................................................................................................... 895.1.2 Logical Type Data...................................................................................................................................................................... 895.1.3 Real and Complex Type Data.....................................................................................................................................................895.1.4 Character Type Data................................................................................................................................................................... 925.1.5 Derived Type Data......................................................................................................................................................................92

5.2 Correct Data Boundaries....................................................................................................................................................................925.3 Precision Conversion......................................................................................................................................................................... 93

5.3.1 Precision Improving....................................................................................................................................................................93

- vii -

5.3.1.1 Precision-Raising Options................................................................................................................................................... 935.3.2 Precision Lowering and Analysis of Errors................................................................................................................................94

5.3.2.1 Defining the Precision-Lowering Function......................................................................................................................... 945.3.2.2 Using the Precision-Lowering Function.............................................................................................................................. 95

Chapter 6 Notes on Programming..........................................................................................................................................976.1 Fortran Versions................................................................................................................................................................................ 976.2 Data....................................................................................................................................................................................................99

6.2.1 COMMON Statement................................................................................................................................................................. 996.2.1.1 Boundaries of Variables in Common Blocks...................................................................................................................... 996.2.1.2 Common Blocks that Have Initial Values........................................................................................................................... 99

6.2.2 EQUIVALENCE Statement..................................................................................................................................................... 1006.2.3 Initialization.............................................................................................................................................................................. 100

6.2.3.1 Character Initialization...................................................................................................................................................... 1006.2.3.2 Initialization with Hexadecimal, Octal, and Binary Constants..........................................................................................1016.2.3.3 Overlapping Initialization.................................................................................................................................................. 101

6.2.4 Using CRAY Pointers...............................................................................................................................................................1016.2.5 SAVE Attribute for Initialized Variables................................................................................................................................. 1026.2.6 Operation of Real or Complex type for Initialization Expression............................................................................................ 102

6.3 Expressions...................................................................................................................................................................................... 1026.3.1 Integer Data Operations............................................................................................................................................................ 1026.3.2 Real Data Operations................................................................................................................................................................ 1036.3.3 Logical Data Operations........................................................................................................................................................... 1036.3.4 Evaluation of Operations.......................................................................................................................................................... 1036.3.5 Order of Evaluation of Data Entities........................................................................................................................................ 1046.3.6 Evaluation of a part of Expression............................................................................................................................................1046.3.7 Definition and Undefinition by Function Reference................................................................................................................ 1046.3.8 Variable enclosed by parenthesis..............................................................................................................................................104

6.4 Assignment Statement..................................................................................................................................................................... 1046.4.1 Assignment Statements with Overlapping Character Positions............................................................................................... 1046.4.2 Allocatable Assignment............................................................................................................................................................105

6.5 Control Statements...........................................................................................................................................................................1056.5.1 GO TO Statement..................................................................................................................................................................... 1056.5.2 Arithmetic IF Statement............................................................................................................................................................1056.5.3 Logical IF Statement.................................................................................................................................................................1066.5.4 DO Statement............................................................................................................................................................................106

6.5.4.1 DO Loop Iteration Count Method..................................................................................................................................... 1066.5.4.2 DO CONCURRENT..........................................................................................................................................................106

6.5.5 CASE Statement....................................................................................................................................................................... 1066.5.6 PAUSE Statement.....................................................................................................................................................................1076.5.7 STOP/ERROR STOP Statement.............................................................................................................................................. 1076.5.8 ALLOCATE/DEALLOCATE Statement.................................................................................................................................1086.5.9 Simple Output List Enclosed in Parentheses............................................................................................................................ 1096.5.10 The Output Statement that starts by TYPE Keyword.............................................................................................................109

6.6 Procedures........................................................................................................................................................................................1096.6.1 Statement Function Reference.................................................................................................................................................. 1106.6.2 An Intrinsic Function that Specifies EXTERNAL Attribute....................................................................................................1106.6.3 REAL and CMPLX Used as Generic Names........................................................................................................................... 1106.6.4 Intrinsic Procedures.................................................................................................................................................................. 1106.6.5 Intrinsic Function Names with the Type Declared................................................................................................................... 1116.6.6 Service Routines....................................................................................................................................................................... 1116.6.7 The returned values of the intrinsic function............................................................................................................................ 111

Chapter 7 Input/output Processing.......................................................................................................................................1127.1 Files..................................................................................................................................................................................................112

7.1.1 Old and New Files.................................................................................................................................................................... 1127.1.2 File Connection.........................................................................................................................................................................112

7.1.2.1 Connection......................................................................................................................................................................... 112

- viii -

7.1.2.2 File disconnection.............................................................................................................................................................. 1127.2 Unit Numbers and File Connection................................................................................................................................................. 113

7.2.1 Priority of Unit and File Connection........................................................................................................................................ 1137.2.2 Environment variables to Connect Units and Files.................................................................................................................. 114

7.3 Records............................................................................................................................................................................................ 1147.3.1 Formatted Fortran Records....................................................................................................................................................... 115

7.3.1.1 Formatted Sequential Record.............................................................................................................................................1167.3.1.2 Formatted Direct Record................................................................................................................................................... 1167.3.1.3 Internal File Record........................................................................................................................................................... 117

7.3.2 Unformatted Fortran Records................................................................................................................................................... 1177.3.2.1 Unformatted Sequential Record.........................................................................................................................................1177.3.2.2 Unformatted Direct Record............................................................................................................................................... 120

7.3.3 List-Directed Fortran Records.................................................................................................................................................. 1207.3.4 Namelist Fortran Records......................................................................................................................................................... 1207.3.5 Endfile Records.........................................................................................................................................................................1207.3.6 Binary Fortran Records.............................................................................................................................................................1207.3.7 STREAM Fortran Records....................................................................................................................................................... 120

7.3.7.1 Formatted Stream Record.................................................................................................................................................. 1207.3.7.2 Unformatted Stream Record.............................................................................................................................................. 121

7.4 OPEN Statement.............................................................................................................................................................................. 1217.4.1 OPEN Statement Specifiers...................................................................................................................................................... 121

7.4.1.1 FILE= Specifier................................................................................................................................................................. 1227.4.1.2 STATUS= Specifier...........................................................................................................................................................1237.4.1.3 RECL= Specifier................................................................................................................................................................1237.4.1.4 ACTION= Specifier...........................................................................................................................................................1247.4.1.5 BLOCKSIZE= Specifier....................................................................................................................................................1257.4.1.6 CONVERT= Specifier.......................................................................................................................................................1257.4.1.7 CARRIAGECONTROL= Specifier.................................................................................................................................. 125

7.4.2 Executing an OPEN Statement on a Connected File................................................................................................................1257.4.3 Input/output of Fortran Records............................................................................................................................................... 1267.4.4 Default Value of ACCESS= Specifier with RECL=Specifier..................................................................................................127

7.5 CLOSE Statement STATUS= Specifier.......................................................................................................................................... 1277.5.1 STATUS= Specifier..................................................................................................................................................................127

7.6 INQUIRE Statement........................................................................................................................................................................ 1287.6.1 INQUIRE Statement Parameters.............................................................................................................................................. 1287.6.2 Difference in Language Version of the INQUIRE statement...................................................................................................133

7.6.2.1 Return value of ACTION= specifier................................................................................................................................. 1337.6.2.2 Char-constant returned by inquiry specifiers of INQUIRE statement...............................................................................1337.6.2.3 Return value of PAD= specifier.........................................................................................................................................1337.6.2.4 Return value of the inquire by file of INQUIREstatement................................................................................................1337.6.2.5 Return value of the inquire by unit of INQUIRE statement.............................................................................................. 133

7.7 Data Transfer Input/output Statements............................................................................................................................................ 1347.7.1 Control Information for Input/output Data Transfer Statements.............................................................................................. 134

7.7.1.1 Input/output UNIT Specifier..............................................................................................................................................1347.7.1.2 Format Specifier................................................................................................................................................................ 1347.7.1.3 IOSTAT= Specifier............................................................................................................................................................1357.7.1.4 Error Branch...................................................................................................................................................................... 1367.7.1.5 End-of-File Branch............................................................................................................................................................ 1367.7.1.6 End-of-Record Branch.......................................................................................................................................................1367.7.1.7 IOMSG= Specifier.............................................................................................................................................................136

7.7.2 Sequential Access Input/output Statements..............................................................................................................................1367.7.2.1 Formatted Sequential Access Input/output Statements..................................................................................................... 1377.7.2.2 Unformatted Sequential Access Input/output Statements................................................................................................. 137

7.7.3 Direct Access Input/output Statements.....................................................................................................................................1377.7.3.1 Direct Access READ Statement........................................................................................................................................ 1387.7.3.2 Direct Access WRITE Statement...................................................................................................................................... 1387.7.3.3 The Error of Data Transfer in a DIRECT Access Input/output.........................................................................................138

- ix -

7.7.4 Namelist Input/output Statements.............................................................................................................................................1397.7.4.1 Namelist READ Statements...............................................................................................................................................139

7.7.4.1.1 Namelist data in Namelist input..................................................................................................................................1417.7.4.2 Namelist WRITE Statements.............................................................................................................................................1427.7.4.3 NAMELIST Input/output Version Differences................................................................................................................. 143

7.7.5 List-Directed Input/output Statements......................................................................................................................................1467.7.5.1 List-Directed READ Statements........................................................................................................................................1467.7.5.2 List-Directed WRITE Statements......................................................................................................................................148

7.7.5.2.1 The Form of Nondelimited Character Constants Produced by List-Directed Output................................................ 1497.7.5.3 List-Directed Input/output Version Differences................................................................................................................ 149

7.7.6 Internal File Input/output Statements....................................................................................................................................... 1507.7.7 Nonadvancing Input/output Statements....................................................................................................................................150

7.8 Combining Input/output Statements................................................................................................................................................ 1507.8.1 Combinations of Input/output Statements Not Allowed...........................................................................................................153

7.8.1.1 Formatted Sequential Access Input/output Statements..................................................................................................... 1537.8.1.2 Unformatted Sequential Access Input/output Statements................................................................................................. 1537.8.1.3 Direct Access Input/output Statements..............................................................................................................................1537.8.1.4 Stream Access Input/output Statements............................................................................................................................ 1547.8.1.5 File Positioning Statements............................................................................................................................................... 154

7.9 File Positioning Statements............................................................................................................................................................. 1547.9.1 BACKSPACE Statement..........................................................................................................................................................1547.9.2 ENDFILE Statement.................................................................................................................................................................1557.9.3 REWIND Statement................................................................................................................................................................. 158

7.10 FLUSH statement.......................................................................................................................................................................... 1587.11 Conversion between IBM370-Format Floating-Point Data and IEEE-Format Floating-Point Data on Input/output...................158

7.11.1 Relationship between Floating-Point Data and Input/output Statements............................................................................... 1587.11.2 Error Processing in the Conversion of Floating-Point Data................................................................................................... 160

7.12 Conversion between Little Endian Data and Big Endian Data on Input/output............................................................................1617.12.1 Relationship between Little Endian Data and Input/output Statements................................................................................. 161

7.13 Standard Files................................................................................................................................................................................ 1627.13.1 Standard Files and Open Modes............................................................................................................................................. 1627.13.2 Restrictions on Input/output Statements for Standard Files................................................................................................... 1627.13.3 Relationship between Input/output Statements and Seekable and Nonseekable Files........................................................... 162

7.14 Restrictions Related to the ACTION= Specifier........................................................................................................................... 1637.15 Number of Significant Digits.........................................................................................................................................................1637.16 I/O Buffer.......................................................................................................................................................................................1637.17 Edit Descriptors............................................................................................................................................................................. 164

7.17.1 Generalized Integer Editing.................................................................................................................................................... 1647.17.2 Exponent Character in Formatted Output...............................................................................................................................1647.17.3 Result of G Editing................................................................................................................................................................. 1647.17.4 Diagnostic Messages for the Character String Edit Descriptor.............................................................................................. 1647.17.5 Effect of the X Edit Descriptor............................................................................................................................................... 1647.17.6 L Edit Descriptor.................................................................................................................................................................... 1657.17.7 D, E, F, G, I, L, B, O, and Z Edit Descriptor without w or d Indicators................................................................................ 1657.17.8 Editing of Special Real-Type Data......................................................................................................................................... 166

7.18 Magnetic Tape File I/O..................................................................................................................................................................1677.18.1 Connection of Magnetic Tape File and Unit Number............................................................................................................ 167

7.18.1.1 Connection by Environment Variable............................................................................................................................. 1677.18.1.2 Connection by FILE Specifier of OPEN Statement........................................................................................................ 167

7.18.2 Note of Magnetic Tape File I/O Function.............................................................................................................................. 1677.19 Conversion between EBCDIC Character Code and ASCII Character Code on Input/output....................................................... 168

7.19.1 Relationship between EBCDIC Character Code Data and Input/output Statements..............................................................1687.20 Asynchronous Input/output Statements......................................................................................................................................... 168

7.20.1 Execution of Asynchronous Input/output Statements............................................................................................................ 1697.20.2 ID= Specifier in Data Transfer Input/output Statements........................................................................................................ 169

Chapter 8 Debugging........................................................................................................................................................... 170

- x -

8.1 Error Processing...............................................................................................................................................................................1708.1.1 Error Monitor............................................................................................................................................................................1708.1.2 Error Subroutines......................................................................................................................................................................175

8.1.2.1 ERRSAV Subroutine......................................................................................................................................................... 1758.1.2.2 ERRSTR Subroutine..........................................................................................................................................................1768.1.2.3 ERRSET Subroutine.......................................................................................................................................................... 1768.1.2.4 ERRTRA Subroutine......................................................................................................................................................... 177

8.1.3 Using the Error Monitor........................................................................................................................................................... 1778.1.3.1 User-Defined Error Subroutines........................................................................................................................................ 1778.1.3.2 Handling of Incorrect Arguments in Intrinsic Functions...................................................................................................1788.1.3.3 User Error Processing........................................................................................................................................................ 178

8.1.4 Input/Output Error Processing.................................................................................................................................................. 1798.1.5 Intrinsic Function Error Processing.......................................................................................................................................... 1828.1.6 Intrinsic Subroutine Error Processing.......................................................................................................................................1978.1.7 Exception Handling Processing................................................................................................................................................ 1998.1.8 Other Error Processing..............................................................................................................................................................200

8.2 Debugging Functions.......................................................................................................................................................................2028.2.1 Debugging Check Functions.....................................................................................................................................................202

8.2.1.1 Checking Argument Validity (ARGCHK)........................................................................................................................ 2048.2.1.1.1 Checking the Number of Arguments.......................................................................................................................... 2048.2.1.1.2 Checking the Argument Types................................................................................................................................... 2048.2.1.1.3 Checking Argument Attributes...................................................................................................................................2048.2.1.1.4 Checking Function Types........................................................................................................................................... 2058.2.1.1.5 Checking Argument Sizes.......................................................................................................................................... 2058.2.1.1.6 Checking Explicit Interface........................................................................................................................................ 2058.2.1.1.7 Checking Rank of Assumed-Shape Array.................................................................................................................. 205

8.2.1.2 Checking Subscript and Substring Values (SUBCHK).....................................................................................................2058.2.1.2.1 Checking Array and Substring References.................................................................................................................2068.2.1.2.2 Checking Array Section References........................................................................................................................... 206

8.2.1.3 Checking References for Undefined Data Items (UNDEF).............................................................................................. 2068.2.1.3.1 Checking for Undefined Data..................................................................................................................................... 2078.2.1.3.2 Checking for Undefined Data Using Invalid Operation Exception............................................................................2078.2.1.3.3 Checking for an Allocatable Variable........................................................................................................................ 2088.2.1.3.4 Checking for an Optional Argument.......................................................................................................................... 208

8.2.1.4 Shape Conformance Check (SHAPECHK).......................................................................................................................2088.2.1.5 Extended Checking of UNDEF (EXTCHK)..................................................................................................................... 2088.2.1.6 Overlapping Dummy Arguments and Extend Undefined CHECK (OVERLAP).............................................................209

8.2.1.6.1 Overlapping Dummy Arguments................................................................................................................................2098.2.1.6.2 Undefined Assumed-size Array with INTENT(OUT)............................................................................................... 2108.2.1.6.3 Undefined Variable of SAVE Attribute......................................................................................................................210

8.2.1.7 Simultaneous OPEN and I/O Recursive Call CHECK (I/O OPEN)................................................................................. 2108.2.1.7.1 Checking Connection of a File to A Unit................................................................................................................... 2118.2.1.7.2 I/O Recursive Call CHECK........................................................................................................................................211

8.2.2 Debugging Programs for Abend............................................................................................................................................... 2118.2.2.1 Causes of Abend................................................................................................................................................................ 2118.2.2.2 Information Generated when an Abend Occurs................................................................................................................ 212

8.2.2.2.1 Information Generated for a General Abend.............................................................................................................. 2128.2.2.2.2 Information of SIGXCPU...........................................................................................................................................2128.2.2.2.3 Information Generated by Run-Time Option -a......................................................................................................... 2128.2.2.2.4 Information when an Abend Occurs Again during Abend Processing.......................................................................212

8.2.3 Hook Function.......................................................................................................................................................................... 2128.2.3.1 User-defined Subroutine Format....................................................................................................................................... 2138.2.3.2 Notes on the Hook Function.............................................................................................................................................. 2138.2.3.3 Calling a User-defined Subroutine from a Specified Location..........................................................................................214

8.2.3.3.1 Notes on Calling from a Specified Location.............................................................................................................. 2158.2.3.4 Calling a User-defined Function at Regular Time Interval............................................................................................... 215

8.2.3.4.1 Notes on Calling by Regular Time interval................................................................................................................ 215

- xi -

8.2.3.5 Calling from Any Location in the Program....................................................................................................................... 215

Chapter 9 Optimization Functions........................................................................................................................................ 2169.1 Overview of Optimization............................................................................................................................................................... 216

9.1.1 Standard Optimization.............................................................................................................................................................. 2169.1.1.1 Elimination of Common Expression..................................................................................................................................2169.1.1.2 Movement of Invariant Expressions.................................................................................................................................. 2169.1.1.3 Reducing the Strength of Operators...................................................................................................................................2179.1.1.4 Loop Unrolling.................................................................................................................................................................. 2179.1.1.5 Loop Blocking................................................................................................................................................................... 2179.1.1.6 Software Pipelining............................................................................................................................................................2189.1.1.7 SIMD................................................................................................................................................................................. 219

9.1.1.7.1 Normal SIMD............................................................................................................................................................. 2199.1.1.7.2 SIMD Extension for Loop Containing IF Construct.................................................................................................. 2199.1.1.7.3 List Vector Conversion...............................................................................................................................................2199.1.1.7.4 SIMD with Redundant Executions for the SIMD Width............................................................................................220

9.1.1.8 Loop Unswitching..............................................................................................................................................................2219.1.1.9 Loop Fission...................................................................................................................................................................... 222

9.1.2 Extended Optimization............................................................................................................................................................. 2229.1.2.1 Inline Expansion................................................................................................................................................................ 2229.1.2.2 Optimization by modifying evaluation methods............................................................................................................... 2239.1.2.3 Advance evaluation of invariant expressions.................................................................................................................... 2249.1.2.4 Loop Striping..................................................................................................................................................................... 2249.1.2.5 XFILL................................................................................................................................................................................ 2259.1.2.6 UXSIMD............................................................................................................................................................................2259.1.2.7 Loop Fission before and after IF constructs in loops(LFI)................................................................................................2269.1.2.8 Loop Versioning................................................................................................................................................................ 2269.1.2.9 CLONE Optimization........................................................................................................................................................ 2279.1.2.10 Link Time Optimization.................................................................................................................................................. 227

9.1.2.10.1 Overview of Link Time Optimization...................................................................................................................... 2279.1.2.10.2 Notes on Link Time Optimization............................................................................................................................ 228

9.2 Notes of wrong erroneous program................................................................................................................................................. 2299.3 Effects of Environmental Change on Optimization.........................................................................................................................2299.4 Effects of Changing the Execution Order........................................................................................................................................2299.5 Branching Target of Assigned GO TO Statements......................................................................................................................... 2299.6 Effect of dummy arguments............................................................................................................................................................ 2309.7 Restrictions on Inline Expansion..................................................................................................................................................... 232

9.7.1 Restrictions on Called User-Defined Procedures..................................................................................................................... 2329.7.2 Restrictions on the Relationship between Calling and Called Program Units......................................................................... 2349.7.3 Restrictions on User-Defined External Procedure Names........................................................................................................2359.7.4 Restrictions on the option......................................................................................................................................................... 235

9.8 Effects of Cray Pointer Variables.................................................................................................................................................... 2359.9 Effects of Compiler Option -Kcommonpad[=N].............................................................................................................................2359.10 Using Optimization control line (OCL).........................................................................................................................................236

9.10.1 Notation rules for optimization control lines..........................................................................................................................2369.10.2 Description position of the optimization control line............................................................................................................. 2369.10.3 Wild Card Specification..........................................................................................................................................................2379.10.4 Optimization Control Specifiers............................................................................................................................................. 237

9.11 Effects of Allocating on Stack....................................................................................................................................................... 2669.12 Effects of Applying Optimization to Change Shape of Array.......................................................................................................2679.13 Effects of generating prefetch on program performance............................................................................................................... 2689.14 The Dimension Shift of the Array Declaration..............................................................................................................................268

9.14.1 Effects of the Dimension Shift of Array Declaration............................................................................................................. 2689.14.2 Restrictions of the Target Variable of the Dimension Shift of the Array Declaration........................................................... 269

9.15 Merging Array Variables............................................................................................................................................................... 2709.15.1 Effects of Merging Array Variables....................................................................................................................................... 2709.15.2 The Restrictions of Merging Local Array Variables.............................................................................................................. 271

- xii -

9.15.3 Restrictions of target variables on common block object array merge...................................................................................2729.16 Multi-Operation Function.............................................................................................................................................................. 272

9.16.1 Calling of Multi-Operation Function......................................................................................................................................2729.16.2 Effects of Compiler Option -Kmfunc=3.................................................................................................................................274

9.17 Software Control of Sector Cache................................................................................................................................................. 2759.17.1 Using Sector Cache.................................................................................................................................................................2759.17.2 How to Control Sector Cache by Software.............................................................................................................................275

9.17.2.1 Software control with optimization control rows............................................................................................................ 2759.17.2.2 Software control with environment variables and optimization control rows.................................................................2769.17.2.3 Behavior when Invalid Value is specified....................................................................................................................... 277

9.18 Incorrectly Specified Optimization Indicators...............................................................................................................................2779.18.1 Example of SIMD(ALIGNED) specifier................................................................................................................................2779.18.2 Example of NOVREC specifier..............................................................................................................................................2789.18.3 Example of CLONE specifier.................................................................................................................................................278

9.19 Attentions for Mixed Language Programming..............................................................................................................................2789.19.1 Address parameters received in the C language .................................................................................................................... 2789.19.2 Pointers received in the C language........................................................................................................................................279

Chapter 10 Fortran Modules.................................................................................................................................................28110.1 Compilation of Modules and Module References......................................................................................................................... 28110.2 Recompiling a Program that Contains a Reference to a Module that Has Changed..................................................................... 28310.3 Restrictions on Modules................................................................................................................................................................ 28310.4 Intrinsic Modules........................................................................................................................................................................... 284

10.4.1 COMPILER_OPTIONS intrinsic module function................................................................................................................28410.4.2 COMPILER_VERSION intrinsic module function............................................................................................................... 284

Chapter 11 Mixed Language Programming..........................................................................................................................28511.1 Summary of Mixed Language Programming................................................................................................................................ 28511.2 Features for Mixed Language Programming................................................................................................................................. 285

11.2.1 Passing Arguments by Value..................................................................................................................................................28611.2.1.1 Passing Arguments by Value in Procedure Interface...................................................................................................... 28611.2.1.2 Passing Arguments by Value in Intrinsic Function......................................................................................................... 286

11.2.2 Getting Arguments by Value.................................................................................................................................................. 28711.2.3 CHANGEENTRY Statement................................................................................................................................................. 28711.2.4 $pragma Specifier................................................................................................................................................................... 28711.2.5 Intrinsic Module ISO_C_BINDING.......................................................................................................................................287

11.2.5.1 Named Constants by Intrinsic Module ISO_C_BINDING............................................................................................. 28711.2.5.2 Types by Intrinsic Module ISO_C_BINDING................................................................................................................29011.2.5.3 Intrinsic Procedures by Intrinsic Module ISO_C_BINDING..........................................................................................290

11.2.6 BIND Statement, BIND Attribute Specifier, Language Binding Specifier and Procedure Language Binding Specifier .....29011.2.6.1 BIND Statement...............................................................................................................................................................29111.2.6.2 BIND Attribute by Attribute Specifier of Type Declaration Statement.......................................................................... 29211.2.6.3 Procedure Language Binding Specifier .......................................................................................................................... 293

11.2.7 VALUE Statement, VALUE Attribute Specifier................................................................................................................... 29311.2.8 Interoperability of Derive type and C language structure type...............................................................................................294

11.3 Rules for Processing Fortran Procedure Names............................................................................................................................ 29411.4 Correspondence of type with C..................................................................................................................................................... 29511.5 Calling Procedures.........................................................................................................................................................................296

11.5.1 Passing Control First to a C Program..................................................................................................................................... 29711.5.2 Calling Standard C Libraries.................................................................................................................................................. 298

11.5.2.1 Method of the Specification of the Procedure Language Binding Specifier .................................................................. 29811.5.2.2 Method of no Specification of the Procedure Language Binding Specifier ................................................................... 299

11.6 Passing and Receiving Function Values........................................................................................................................................29911.6.1 Passing Fortran Function Values to C without the Procedure Language Binding Specifier .................................................29911.6.2 Receiving Function Values from C without the Procedure Language Binding Specifier...................................................... 302

11.7 Passing and Getting Arguments.....................................................................................................................................................30311.7.1 Passing Arguments from Fortran with the Procedure Language Binding Specifier to C.......................................................30311.7.2 Passing Arguments from C to Fortran without the Procedure Language Binding Specifier..................................................303

- xiii -

11.8 Passing Using External Variables..................................................................................................................................................30411.8.1 Passing External Variables with BIND Attribute................................................................................................................... 30511.8.2 Passing External Variables without BIND attribute............................................................................................................... 305

11.9 Passing Using Files........................................................................................................................................................................30611.10 Array Storage Sequence...............................................................................................................................................................30611.11 Passing a Character String to a C Program..................................................................................................................................30711.12 Return Value of Executable Program.......................................................................................................................................... 30811.13 Notes of the Compiler Option -Az...............................................................................................................................................30811.14 Restrictions on Linking with C Language Programs...................................................................................................................308

Chapter 12 Multiprocessing..................................................................................................................................................31012.1 Overview of Multiprocessing........................................................................................................................................................ 310

12.1.1 Parallel processing.................................................................................................................................................................. 31012.1.2 Effect of Multiprocessing....................................................................................................................................................... 31012.1.3 Requirements for Effective Multiprocessing..........................................................................................................................31112.1.4 Features of Parallel Processing in this Compiler.................................................................................................................... 311

12.2 Automatic Parallelization.............................................................................................................................................................. 31112.2.1 Compilation and Execution.................................................................................................................................................... 311

12.2.1.1 Compilation..................................................................................................................................................................... 31112.2.1.1.1 Compiler Option for Automatic Parallelization........................................................................................................311

12.2.1.2 Execution Process............................................................................................................................................................ 31212.2.1.2.1 Environment Variable PARALLEL......................................................................................................................... 31212.2.1.2.2 Number of Threads................................................................................................................................................... 31312.2.1.2.3 Environment variables THREAD_STACK_SIZE and OMP_STACKSIZE........................................................... 31312.2.1.2.4 Stack Size on Execution........................................................................................................................................... 31312.2.1.2.5 The Process of Waiting Threads...............................................................................................................................31412.2.1.2.6 Notes when setting time limits..................................................................................................................................31412.2.1.2.7 Notes when using service routines............................................................................................................................31412.2.1.2.8 CPU Binding for Thread...........................................................................................................................................314

12.2.1.3 Example of Compilation and Execution..........................................................................................................................31612.2.2 Tuning of parallelized programs.............................................................................................................................................31612.2.3 Details of Automatic Parallelization.......................................................................................................................................316

12.2.3.1 Automatic Parallelization................................................................................................................................................ 31612.2.3.1.1 Targets for Automatic Parallelization.......................................................................................................................31612.2.3.1.2 What is Loop Slicing?...............................................................................................................................................31612.2.3.1.3 Array Operations and Automatic Parallelization......................................................................................................31712.2.3.1.4 Automatic Loop Slicing by the Compiler.................................................................................................................31712.2.3.1.5 Loop Interchange and Automatic Loop Slicing........................................................................................................31712.2.3.1.6 Loop Distribution and Automatic Loop Slicing....................................................................................................... 31812.2.3.1.7 Loop Fusion and Automatic Loop Slicing................................................................................................................31912.2.3.1.8 Loop Reduction.........................................................................................................................................................31912.2.3.1.9 Restrictions on Loop Slicing.....................................................................................................................................32012.2.3.1.10 Displaying the State of Automatic Parallelization..................................................................................................32212.2.3.1.11 Extension of Parallel Region.................................................................................................................................. 32212.2.3.1.12 Block Distribution and Cyclic Distribution............................................................................................................ 323

12.2.3.2 Optimization Control Line...............................................................................................................................................32412.2.3.2.1 Notation rules for optimization control lines............................................................................................................32412.2.3.2.2 Description position of the optimization control line............................................................................................... 32412.2.3.2.3 Automatic Parallelization and Optimization Control Specifiers.............................................................................. 32412.2.3.2.4 Optimization Control Specifiers for Automatic Parallelization............................................................................... 32412.2.3.2.5 Wild Card Specification............................................................................................................................................338

12.2.3.3 Notes on Automatic Parallelization................................................................................................................................. 33912.2.3.3.1 Caution when specifying compile-time option -Kparallel,instance=N.................................................................... 33912.2.3.3.2 Nested Parallel Processing........................................................................................................................................33912.2.3.3.3 Caution when specifying compile-time option -Kparallel,reduction........................................................................34012.2.3.3.4 Notes on Using Optimization Control Line..............................................................................................................34012.2.3.3.5 I/O Statement and Intrinsic Procedure Call on Parallel Processing..........................................................................341

- xiv -

12.2.4 Linkage with Other Multi-Thread Programs.......................................................................................................................... 34212.3 Parallelization by OpenMP Specifications.................................................................................................................................... 343

12.3.1 Compilation and Execution.................................................................................................................................................... 34312.3.1.1 Compilation..................................................................................................................................................................... 343

12.3.1.1.1 Compiler Options to Compile OpenMP Programs................................................................................................... 34312.3.1.1.2 Displaying Optimization Information of OpenMP Program.................................................................................... 345

12.3.1.2 Execution Process............................................................................................................................................................ 34512.3.1.2.1 Environment Variable at Execution..........................................................................................................................34512.3.1.2.2 Environment Variable for OpenMP Specifications.................................................................................................. 34712.3.1.2.3 Notes during Execution............................................................................................................................................ 34712.3.1.2.4 Output from Multiple Threads..................................................................................................................................350

12.3.2 Implementation-Dependent Specifications.............................................................................................................................35012.3.3 Clarifying OPENMP Specifications and restrictions .............................................................................................................353

12.3.3.1 Specifying Label with ASSIGN Statement and Assigned GO TO Statement.................................................................35312.3.3.2 Additional Functions and Operators in ATOMIC Directive and REDUCTION Clause................................................ 35312.3.3.3 Notes on Using THREADPRIVATE.............................................................................................................................. 35312.3.3.4 Inline Expansion.............................................................................................................................................................. 35412.3.3.5 Internal Procedure Calling from Parallel Region............................................................................................................ 35412.3.3.6 Polymorphic Variable...................................................................................................................................................... 35412.3.3.7 BIND attribute................................................................................................................................................................. 35412.3.3.8 OPTIONAL attribute....................................................................................................................................................... 35412.3.3.9 ASSOCIATE NAME.......................................................................................................................................................35412.3.3.10 Selector.......................................................................................................................................................................... 35412.3.3.11 Derived type variable with length type parameter.........................................................................................................355

12.3.4 Notes on Creating Program.....................................................................................................................................................35512.3.4.1 Implementing PARALLEL Region and Explicit TASK Region.....................................................................................35512.3.4.2 Automatic Parallelization for OpenMP Programs........................................................................................................... 355

12.3.5 Linkage with Other Multi-Thread Programs.......................................................................................................................... 35512.3.6 Debug OpenMP Programs...................................................................................................................................................... 356

12.3.6.1 Checking Features for Debugging................................................................................................................................... 35712.4 I/O Buffer Parallel Transfer...........................................................................................................................................................357

12.4.1 Compilation and Execution.................................................................................................................................................... 35712.4.1.1 Compilation..................................................................................................................................................................... 35712.4.1.2 Execution Process............................................................................................................................................................ 35712.4.1.3 Conditions to Perform I/O Buffer Parallel Transfer........................................................................................................ 35712.4.1.4 Compilation execution example...................................................................................................................................... 357

12.4.2 Notes on I/O Buffer Parallel Transfer.....................................................................................................................................358

Appendix A Limits and Restrictions......................................................................................................................................359A.1 Nesting Level of Functions, Array Sections, Array Elements, and Substring References.............................................................359A.2 DO, CASE, IF, SELECT TYPE, BLOCK, CRITICAL, and ASSOCIATE Constructs................................................................ 359A.3 Implied DO-Loops.......................................................................................................................................................................... 359A.4 INCLUDE Line...............................................................................................................................................................................359A.5 Array Declarations.......................................................................................................................................................................... 360

A.5.1 Restriction for All Array Declarations.....................................................................................................................................360A.5.2 Restriction for Array Declaration of Zero-Sized Array...........................................................................................................360

A.6 Array Section.................................................................................................................................................................................. 361A.7 Distance between a Branch Instruction and the Branch Destination Instruction............................................................................361

Appendix B Printing Control Command................................................................................................................................362B.1 Command Format............................................................................................................................................................................362B.2 Example of fot Command and fotpx Command............................................................................................................................. 362B.3 Return Values of fot Command and fotpx Command.................................................................................................................... 363B.4 Diagnostic Messages of fot and fotpx.............................................................................................................................................363

Appendix C Endian Convert Command................................................................................................................................364C.1 Command Format............................................................................................................................................................................364C.2 Return Value of fcvendian(1) and fcvendianpx(1)......................................................................................................................... 364

- xv -

C.3 Diagnostic Messages of fcvendian(1) and fcvendianpx(1).............................................................................................................364C.4 Note................................................................................................................................................................................................. 365

Appendix D Job Operation Software.................................................................................................................................... 366D.1 Management of the CPUs resource.................................................................................................................................................366

D.1.1 Number of CPUs that can be used on the Job..........................................................................................................................366D.1.2 FLIB_USE_ALLCPU..............................................................................................................................................................366D.1.3 FLIB_USE_CPURESOURCE.................................................................................................................................................367

D.2 CPU Binding...................................................................................................................................................................................367D.2.1 FLIB_CPUBIND..................................................................................................................................................................... 367

D.3 On-chip Hardware Barrier.............................................................................................................................................................. 368D.3.1 Compilation..............................................................................................................................................................................368D.3.2 Execution................................................................................................................................................................................. 368D.3.3 FLIB_CNTL_BARRIER_ERR............................................................................................................................................... 368D.3.4 FLIB_NOHARDBARRIER.................................................................................................................................................... 368D.3.5 Notes........................................................................................................................................................................................ 368D.3.6 Usage example on the job........................................................................................................................................................ 369

D.4 Sector-cache....................................................................................................................................................................................369D.4.1 Notes in execution....................................................................................................................................................................369

D.5 Large-Page Function.......................................................................................................................................................................370

Appendix E Preprocessor.....................................................................................................................................................371E.1 C language preprocessor................................................................................................................................................................. 371E.2 Fortran preprocessor........................................................................................................................................................................371

E.2.1 Continuation line of Fortran..................................................................................................................................................... 371E.2.2 Comment line of Fortran.......................................................................................................................................................... 371E.2.3 Comment of Fortran and comment of C language................................................................................................................... 372E.2.4 Macro substitution in Fortran comment construct................................................................................................................... 372E.2.5 Treatment of the characters specified after column 72 in fixed source form...........................................................................372

Appendix F Notes on Migration from FX10 System to FX100 System.................................................................................374F.1 Error check in OpenMP standard at compilation time.................................................................................................................... 374F.2 Error check in Fortran 2003 standard at compilation time..............................................................................................................374F.3 Enhancement intrinsic procedures and intrinsic module.................................................................................................................374F.4 The option which the compiler option -Kuxsimd needs is changed................................................................................................374F.5 The function for canceling the program compilation that is forecasted to take a long time........................................................... 374F.6 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|

openmp_noordered_reduction}................................................................................................................................................ 375F.7 Abolition of compilation options and the optimization control specifiers...................................................................................... 375F.8 The -Komitfp option is induced by -Kfast, and maintenance of the trace back information.......................................................... 377F.9 Change of value of macro along with the compiler version up.......................................................................................................377F.10 Change of the name of temporary object file and the directory to store....................................................................................... 377F.11 Change the default value of the -Ksimd[=level] option................................................................................................................ 377F.12 Specification change of the optimization control specifiers SIMD and UXSIMD....................................................................... 378

Appendix G Compatibility Information (FX10 System)......................................................................................................... 379G.1 Migrating to V2.0L20(Generation Number:12)............................................................................................................................. 379

G.1.1 Error check in OpenMP standard at compilation time.............................................................................................................379G.1.2 Error check in Fortran 2003 standard at compilation time...................................................................................................... 379

G.2 Migrating to V2.0L10(Generation Number:11)............................................................................................................................. 381G.2.1 The function for canceling the program compilation that is forecasted to take a long time....................................................381G.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|

openmp_noordered_reduction}........................................................................................................................................ 381G.3 Migrating to V2.0L10(Generation Number:10)............................................................................................................................. 382

G.3.1 The -Komitfp option is induced by -Kfast, and maintenance of the trace back information.................................................. 382G.3.2 Change of value of macro along with the compiler version up............................................................................................... 382G.3.3 Change of the name of temporary object file and the directory to store..................................................................................382

G.4 Migrating to V1.0L30(Generation Number:09)............................................................................................................................. 383

- xvi -

G.4.1 Change of values of macro and named constant along with the support of OpenMP API version 3.1 specifications............383G.4.2 Enhancement Fortran 2008 intrinsic procedures and intrinsic modules..................................................................................383G.4.3 Bound remapping list and pointer target of two rank or more in pointer assignment statement.............................................385G.4.4 Non pure final subroutine referred from pure procedure.........................................................................................................385G.4.5 Array argument that is not CONTIGUOUS in intrinsic module function C_LOC................................................................. 386G.4.6 Runtime message (jwe1007i-s) change................................................................................................................................... 386

G.5 Migrating to V1.0L20(Generation Number:07)............................................................................................................................. 386G.5.1 Error message output when specifying parallelization options are omitted at linking............................................................ 386

G.6 Migrating to V1.0L20(Generation Number:06)............................................................................................................................. 387G.6.1 Message output for unrecognized compiler option..................................................................................................................387

G.7 Migrating to V1.0L20..................................................................................................................................................................... 388G.7.1 Runtime message (jwe0220i-e) change................................................................................................................................... 388G.7.2 Changes to the specification when SOURCE= specifier appears and an allocatable component is "allocated" in ALLOCATE

statement...........................................................................................................................................................................388G.7.3 Support Fortran 2008 intrinsic procedures.............................................................................................................................. 389G.7.4 Change when specifying the -O option before specifying the -g option................................................................................. 389

Appendix H Compatibility Information (FX100 System)....................................................................................................... 391H.1 Migrating to V2.0L20(Generation Number:03)............................................................................................................................. 391

H.1.1 Error check in OpenMP standard at compilation time.............................................................................................................391H.1.2 Error check in Fortran 2003 standard at compilation time...................................................................................................... 391H.1.3 Enhancement intrinsic procedures and intrinsic module......................................................................................................... 391H.1.4 Change of the compilation message when the compiler option -K{FLTLD|NOFLTLD} is specified...................................393

H.2 Migrating to V2.0L10(Generation Number:02)............................................................................................................................. 393H.2.1 The function for canceling the program compilation that is forecasted to take a long time....................................................393H.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|

openmp_noordered_reduction}........................................................................................................................................ 393

Index.....................................................................................................................................................................................394

- xvii -

Chapter 1 Overview of the Fortran SystemThis chapter provides an overview of the Fortran system (or simply "the system") and illustrates the most common ways to use the systemto develop and run a Fortran program.

1.1 Configuration of the Fortran SystemThe Fujitsu Fortran system runs under the Operating System based on Linux.

The system consists of the following components. There are two types of compile commands, one is the cross compile command that isused on the login node, and the other is the own compile command that is used on the compute node.

- The Fortran driver frtpx (cross compiler command) or frt (own compiler command), calls the Fortran compiler, the preprocessor, theassembler, and the linker.

The frtpx is used as the name of the driver at the following. Replace the frtpx as the frt when you use the own compiler.

- The Fortran compiler creates object programs from Fortran source programs.

- The Fortran library provides commonly used start-up routines, input/output facilities, and intrinsic procedures.

- Printing control command fot/fotpx and endian convert command fcvendian/fcvendianpx are included in the system.

- Online manual pages, accessed by the man command, provide information about frtpx(1), frt(1), fot(1), fotpx(1), fcvendian(1),fcvendianpx(1), intrinsic(3) and service(3).

The following is cross compiler command and own compiler command. Please read frtpx in a different way as frt when you use owncompiler.

Component Cross Compiler Own Compiler

Fortran driver frtpx frt

Printing control command fotpx fot

Endian convert command fcvendianpx fcvendian

1.2 How to use Fortran System

1.2.1 PreparationTo use the system, following environmental variables must be set correctly.

Environment variable Value

PATH /opt/FJSVmxlang/bin

LD_LIBRARY_PATH /opt/FJSVmxlang/lib64

MANPATH /opt/FJSVmxlang/man

Examples:

$ PATH=/opt/FJSVmxlang/bin:$PATH ; export PATH$ LD_LIBRARY_PATH=/opt/FJSVmxlang/lib64:$LD_LIBRARY_PATH ; export LD_LIBRARY_PATH$ MANPATH=/opt/FJSVmxlang/man:$MANPATH ; export MANPATH

1.2.2 Compilation and LinkingThe frtpx command is used for compilation and linking.

The simplest case is to compile a Fortran program, stored in a file such as prog.f95. The command

- 1 -

$ frtpx prog.f95

compiles and links the program, producing an executable file named a.out. Specifying -c option causes only object file to be made, withoutlinking.

$ frtpx -c prog.f95

1.2.3 File NamesThe frtpx command may specify source files, object files, or assembly language files. Source files may havea .f90, .f95, .f, .for, .f03, .F90, .F95, .F, .FOR or .F03 suffix. If it is .f90, .f95 or .f03, the statements are assumed to be in free source form;if it is .f or .for, the statements are assumed to be in fixed source form; in either case, this assumption may be overridden with the -Freeor -Fixed compiler option. If the suffix is .F90, .F95, .F, .FOR or .F03, the preprocessor is invoked before the compiler and the assumedsource form is fixed and free, respectively.

Assembly language files have a .s suffix; they are not processed by the compiler but passed on to the assembler. An assembly languagefile is produced when the -S option is used.

Object files have a .o suffix. Such files are not processed by the compiler, but passed on to the linker.

1.2.4 Some Compilation OptionsThere are many compilation options; they are described in detail in Section "2.2 Compiler Options". Some of the most important andcommonly used options are:

-c Produce object files only and do not link to create an executable.

-fw Output only warning and serious error messages (no informative messages).

-fs Output only serious error messages.

-fmsg_num Suppress particular error message specified by msg_num.

-Nmaxserious=maxnum Stop the compilation if s-level (serious) error detected more maxnum.

-oexe_file The executable is named exe_file instead of a.out. If the -c option is alsospecified, the object file is named exe_file.

-Fixed The program is in fixed source form.

-Free The program is in free source form.

-H[a|e|f|o|s|u|x] Check for argument mismatches (a), shape conformance (e), simultaneousOPEN and I/O recursive call (f), overlapping dummy arguments and extendundefined (o), subscript and substring values out of bounds (s), undefinedvariables (u), or module, common, and pointer undefined check (x). Thesemay be combined: -Haesu or -Hsu.

-Mdirectory Names a directory to hold module information files

-Idirectory Names a directory to search for include and module information files.

-O[{0|1|2|3}] Specifies the optimization level. The default is the -O2.

-Kfast Sets the many optimization parameters to reasonable values to achieveoptimized performance

-X6 Compiles Fortran source program as FORTRAN66 source.

-X7 Compiles Fortran source program as FORTRAN77 source.

-X9 Compiles Fortran source program as Fortran 95 source.

-X03 Compiles Fortran source program as Fortran 2003 source.

1.2.5 ExecutionYou can run an executable program created by frtpx command. An example of executing a.out file follows:

- 2 -

$ ./a.out

You can specify instructions to Fortran library on invoking an executable file (see "Chapter 3 Executing Fortran Programs").

1.2.6 DebuggingThere are two methods of debugging Fortran programs; one is batch debug function in this Fortran system, another is source-levelinteractive debug function.

See Section "8.2 Debugging Functions" for batch debug function.

See the "Debugger User's Guide" for source-level interactive debug function.

1.2.7 TuningThere is a sampling function provided by the Profiler for tuning of Fortran programs.

See the "Profiler User's Guide" for the Profiler.

- 3 -

Chapter 2 Compiling and Linking Fortran ProgramsThis chapter tells you how to compile and link Fortran programs.

2.1 Compile CommandUse the compile command frtpx to compile and prepare Fortran programs for execution. The frtpx command analyzes its arguments. Itthen calls the cpp command for C preprocessing, the fpp command for Fortran preprocessing, Fortran compiler, the as command for theassembler, and the ld command for the linker, as necessary.

2.1.1 Compile Command FormatSpecify the filename list and option list as arguments for the Fortran compiler, the as command, and the ld command. See Section "2.2Compiler Options" for information on Fortran compiler options. For information on the as, and ld commands, use the man command torefer to the online manual.

The format of the frtpx command is as follows:

Command Operand

frtpx [_option-list]_file-name-list

_: At least one blank is required.

Notes:

1. The items in the option-list and file-name-list can be specified in any order. For example, you may specify the option-list after thefile-name-list or mix option-list and file-name-list specifications.

2. At least one file name must be specified in the file-name-list, unless the entire operand is -V.

3. See Section "2.1.2 Compile Command Input File" for details of file-name-list.

2.1.2 Compile Command Input FileThe following table lists the suffixes allowed for frtpx(1) file names.

File type File suffix Program

Fortran source program

.f

Fortran compiler

.for

.f90

.f95

.f03

Preprocessing directive Fortran source program(*1)

.F

Preprocessor followed by Fortrancompiler

.FOR

.F90

.F95

.F03

Assembler source program .s Assembler

Object program .o Linker

Other than above Other than above Linker

*1: Preprocessing directives may be included in Fortran source programs. They have the same form as directives for the C language;each line begins with the symbol "#".

- 4 -

2.2 Compiler OptionsCompiler options may be specified as option arguments of frtpx or by the compilation profile file or by the compiler directive lines or bythe environment variable FORT90CPX or FORT90C.

See Section "2.3 Compile Command Environment Variable" for information about the environment variable FORT90CPX or FORT90C.

See Section "2.4 Compilation Profile File" for information about the compilation profile file.

See Section "2.5 Compiler Directive Lines" for information about the compiler directive lines.

This section explains the format and function of compiler options.

Each of the system components may be affected by the options.

The following table shows the relation between a file suffix and some compiler options.

File suffix Option

.f

-X03 -Fixed (*1).for

.F

.FOR

.f90

-X9 -Free.f95

.F90

.F95

.f03-X03 -Free

.F03

*1: If the -Xd7 option is specified, activates -X7 -Fixed options instead.

Compiler Options

Table 2.1 [Options for preprocessor]

Function overview Option name

The -Ccpp option specifies to use the C preprocessor. -Ccpp

The -Cfpp option specifies to use the Fortran preprocessor. -Cfpp

The -Cpp option specifies to use preprocessor. -Cpp

The -D option specifies to replace name by value. -Dname[=tokens]

The -P option specifies to make preprocessor store results in thecurrent directory.

-P

The -E option specifies to perform only preprocessing to thespecified source file.

The result of preprocessing is output to the standard-output.

-E

The -U option specifies to invalidate the definition of name. -Uname

Table 2.2 [Options for source program]


The -AU option specifies to treat characters in case sensitivemanner.

-AU

The -Fixed option specifies to compile a Fortran source programas a fixed-form source.

-Fixed

- 5 -


The -Free option specifies to compile a Fortran source program asa free-form source.

-Free

The -Fwide option specifies to compile a Fortran source programas a fixed-form source whose length of all lines is 255.

-Fwide

The -Xd7 option specifies that -X7 option is the default oflanguage-version option when the suffix of the Fortran source fileis .f, .for, .F or .FOR.

-Xd7

Table 2.3 [Options for Fortran standards and language specifications]


The -AE option specifies not to treat special characters such asbackslash (\) as control-characters.

-AE

The -Ae option specifies to treat special characters such asbackslash (\) as control-characters.

-Ae

The -AT option indicates that it does not apply implicit typing andspecifies the null mapping for all the letters.

-AT

The -Ap option specifies to treat the default access permission ofmodules as PRIVATE.

-Ap

The -Aw option specifies to compile IACHAR, ACHAR, IBITSand ISHFTC functions as intrinsic procedures.

-Aw

The -Ay option specifies to assume type of an unsigned realconstant on the right side of an assignment statement as type of avariable on the left side.

-Ay

The -Az option specifies to append a null character (\0) to the endof each string constant argument of an external procedure.

-Az

These options specify whether or not to enable external namewhich appears only in EXTERNAL statements.

-N{allextput | noallextput}

These options specify whether or not to process the allocatableassignment of the Fortran 2003 standard.

-N{alloc_assign | noalloc_assign}

These options specify whether or not to enable COARRAYspecifications.

-N{ coarray | nocoarray }

These options specify whether or not to copy constant actualargument of procedure.

-N{copyarg | nocopyarg}

These options specify whether or not to process the characterassignment statements in Fortran standard.

-N{f90move | nof90move}

These options specify whether or not to assume MALLOC andFREE as intrinsic procedures.

-N{mallocfree | nomallocfree}

These options specify whether or not to compile particularfunctions as intrinsic functions.

-N{obsfun | noobsfun}

These options control an initial state for a list item with theALLOCATABLE attribute in a private clause of OpenMP.

-N{ privatealloc | noprivatealloc }

These options specify whether or not to add RECURSIVE keywordin each SUBROUTINE and FUNCTION statement.

-N{recursive | norecursive}

These options specify whether or not to add SAVE statement ineach program unit except main program.

-N{save | nosave}

- 6 -


These options specify whether or not to automatically set zerovalue by the procedure entrance and ALLOCATE statement ineach variable.

-N{ setvalue[=set_arg] | nosetvalue }

set_arg: { { heap | noheap } | { stack | nostack } |{ scalar | noscalar } | { array | noarray } | { struct |nostruct } }

These options specify version of the language. -X{6 | 7 | 9 | 03}

Table 2.4 [Options for degree of accuracy]


These options specify to improve the precision of the variables.

-A{d | i | q}

-C{ cdII8 | cI4I8 | cdLL8 | cL4L8 | cdRR8 | cR4R8| cd4d8 | ca4a8 | cdDR16 | cR8R16 }

The -Cl option specifies to verify whether or not precision ofrounding errors is in specified range.

-Cl

0 <= l <= 15

Table 2.5 [Options for data allocation]


The -AA option specifies not to align data in common blocks. -AA

These options specify whether or not to free allocatable array. -N{freealloc | nofreealloc}

These options specify whether or not to allocate data to write inhibitarea.

-N{use_rodata | nouse_rodata}

Table 2.6 [Options for math functions]


These options specify whether or not to inline intrinsic functions. -K{ilfunc | noilfunc}

These options specify whether or not to use multi-operationfunction.

-K{mfunc[=level] | nomfunc}

level: {1 | 2 | 3}

These options specify whether or not to create executable programwhich the Fortran intrinsic function library is linked in statically.

-K{ static_fjlib | nostatic_fjlib }

The -SSL2 option specifies to link SSL-II and sequential BLAS/LAPACK libraries.

-SSL2

The -SSL2 option specifies to link SSL-II and parallelized BLAS/LAPACK libraries.

-SSL2BLAMP

Table 2.7 [Options for performance and optimization]


The -O option specifies the optimization level. -O{ 0 | 1 | 2 | 3 }

The -x option specifies to apply inline expansion foruser-defined procedures.

-x{ - | proc_name[,proc_name]... | stmt_no | dat_szK |dir=dir_name[,dir=dir_name]... | 0 }

The -Kadr44 option specifies to create object files inwhich absolute address is expressed as 44-bit data.

-Kadr44

The -Kadr64 option specifies to create object files inwhich absolute address is expressed as 64-bit data.

-Kadr64

The -Knoalias option specifies to performoptimization under the assumption that pointers in thesource code do not refer to the same memory area as

-Knoalias[=spec]

spec: 's'

- 7 -


other pointers or variables. 's' can be specified inspec.

These options specify to pad arrays.

-Karraypad_const[=N]

1 <= N <= 2147483647

-Karraypad_expr=N

1 <= N <= 2147483647

These options specify to merge arrays.

-Karray_merge

-Karray_merge_common[=name]

-Karray_merge_local

-Karray_merge_local_size=N

2 <= N <= 2147483647

These options specify to shift dimensions of arrays.

-Karray_subscript

-Karray_subscript element=N

2 <= N <= 2147483647

-Karray_subscript_elementlast=N

2 <= N <= 2147483647

-Karray_subscript_rank=M

2 <= M <= 30

-Karray_subscript_variable='ary_nm(rank,rank[,rank]...)'

1 <= rank <=30

The -Karray_transform option is equivalent to -Karraypad_const,array_merge, array_subscript.

-Karray_transform

These options specify whether or not to allocate localvariable on the stack.

-K{auto | noauto}

These options specify whether to allocate automaticdata on the stack or on the heap.

-K{autoobjstack | noautoobjstack}

These options specify whether to allocate the resultarea of user-defined array function statically in thecaller program unit or dynamically in execution.

-K{calleralloc[=level] | nocalleralloc}

level: {1 | 2}

The -Kcommonpad option specifies to pad array incommon blocks or module.

-Kcommonpad[=N]

4 <= N <= 2147483644

These options specify whether or not to assume thatthe particular data is aligned on the eight-byteboundary.

-K{dalign | nodalign}

These options specify whether or not to performoptimization that modifies how operations areevaluated for object programs.

-K{eval | noeval}

The -Kfast option specifies to create an object programwhich runs at high speed on the target machine.

-Kfast

These options specify whether or not to create anexecutable program which uses floating-pointexception disable mode.

-K{ fed | nofed }

- 8 -


These options specify whether or not the program mayaccess floating point rounding mode, and also specifywhether or not the rounding mode is the default.

-K{ fenv_access | nofenv_access }

These options specify whether or not to applyoptimizations which use non-faulting loadinstructions.

-K{FLTLD | NOFLTLD}

These options direct whether or not to performoptimization using "Floating-Point Multiply-Add/Subtract" floating point operation instructions onobject programs.

-K{fp_contract | nofp_contract}

These options specify whether or not to convert afloating point division or a SQRT function intoreciprocal approximation instructions.

-K{fp_relaxed | nofp_relaxed}

These options specify whether or not to applyoptimizations which simplify floating-pointoperations.

-K{fsimple | nofsimple}

These options specify whether to generate the objectfile using instruction set for either HPC-ACE or HPC-ACE2.

-K{HPC_ACE | HPC_ACE2}

These options direct whether the optimization whichuses the intent is performed in the program unit ofprocedure call which includes the actual argumentassociated with the dummy argument to whichINTENT attribute is specified.

-K{intentopt | nointentopt}

These options specify whether or not to apply loopblocking optimization.

-K{loop_blocking[=N] | loop_noblocking}

2 <= N <= 10000

These options specify whether or not to perform loopfission:

the optimization which divides a loop into pluralloops.

-K{ loop_fission | loop_nofission }

These options specify whether or not to perform theloop fission optimization as much as possible beforeand after IF constructs in loops.

-K{ loop_fission_if | loop_nofission_if }

These options specify whether or not to fuse loops intoone loop.

-K{loop_fusion | loop_nofusion}

These options specify whether or not to exchangeloops.

-K{ loop_interchange | loop_nointerchange }

These options specify whether or not to performoptimization which divides loops and partially usesSIMD extensions.

-K{ loop_part_simd | loop_nopart_simd }

These options specify whether or not to performoptimization by the loop versioning.

-K{ loop_versioning | loop_noversioning }

These options specify whether or not to perform linktime optimization.

-K{lto|nolto}

These options specify whether or not to applyoptimizations which use the Non-Faulting mode.

-K{ nf| nonf }

These options specify whether or not to let FPU benonstandard mode.

-K{ns | nons}

- 9 -


These options specify whether or not to make OCLs(Optimization Control Lines) effective.

-K{ocl | noocl}

-Kvppocl

These options specify whether or not to keep the framepointer in a register from procedure calling.

-K{omitfp | noomitfp}

The -Koptions specifies to treat "!options" statementsas compiler directive lines.

-Koptions

These options specify to generate position-independent code (PIC).

-K{pic | PIC}

These options specify whether or not to evaluateinvariant expressions in advance.

-K{preex | nopreex}

These options specify how to generate prefetchinstructions.

-Kprefetch_cache_level={1 | 2 | all}

-K{prefetch_conditional | prefetch_noconditional}

-K{prefetch_indirect | prefetch_noindirect}

-K{prefetch_infer | prefetch_noinfer}

-Kprefetch_iteration=N

1 <= N <= 10000

-Kprefetch_iteration_L2=N

1 <= N <= 10000

-Kprefetch_line=M

1 <= M <= 100

-Kprefetch_line_L2=M

1 <= M <= 100

-K{ prefetch_sequential[=kind] | prefetch_nosequential }

kind: {auto | soft}

-K{prefetch_stride | prefetch_nostride}

-Knoprefetch

-K{ prefetch_strong | prefetch_nostrong }

-K{ prefetch_strong_L2 | prefetch_nostrong_L2 }

These options specify whether some optimizations forshort loops are applied to all innermost loops in theprogram.

-K{shortloop=N | noshortloop}

2 <= N <= 10

These options specify whether or not to use SIMDextensions.

-K{simd[=level] | nosimd}

level: {1 | 2 | auto}

These options specify to determine whether to useSIMD indirect instructions or scalar instructions.

-K{simd_separate_stride | simd_noseparate_stride}

These options specify whether or not to perform thestriping optimization.

-K{ striping[=N] | nostriping }

2 <= N <= 100

These options specify whether or not to apply softwarepipelining.

-K{swp | noswp}

This option specifies to perform software pipeliningfor more loops by easing its applicable condition.

-Kswp_strong

- 10 -


These options specify whether or not to allocatetemporary array data object on the stack.

-K{temparraystack | notemparraystack}

These options specify whether or not to apply loopunrolling .

-K{unroll[=N] | nounroll}

2 <= N <= 100

These options specify whether or not to use SIMDextensions by UXSIMD optimization.

-K{uxsimd | nouxsimd}

These options specify whether or not to generateXFILL instructions.

-K{ XFILL[=N] | NOXFILL }

1 <= N <= 100

Table 2.8 [Options for thread parallelization]


These options specify whether or not to localize arrays in loops. -K{array_private | noarray_private}

These options specify whether or not to select a loop to beparallelized considering the number of threads when both an innerloop and an outer loop are candidates in a nested loop.

-K{dynamic_iteration | nodynamic_iteration}

The -Kindependent option specifies that the procedure namedproc_name behave just the same regardless whether DO loops areparallelized or not.

-Kindependent=proc_name

The -Kinstance option specifies to create a parallelized objectprogram executed in N threads.

-Kinstance=N

2 <= N <= 512

These options specify whether or not to perform automaticparallelization that requires dividing loops.

-K{ loop_part_parallel | loop_nopart_parallel }

These options specify whether or not to enable directives ofOpenMP standard.

-K{openmp | noopenmp}

These options direct whether or not to promote optimization byassuming that array elements of the chunk size have no datadependency over iteration.

-K{ openmp_assume_norecurrence |openmp_noassume_norecurrence }

These options specify whether or not to fix the operation order ofthe reduction operations same as in the numerical order of thethreads at the end of the region for which the reduction clause ofOpenMP was specified.

-K{ openmp_ordered_reduction |openmp_noordered_reduction }

These options specify whether or not to allocate the threadprivatevariable on the Thread-Local Storage.

-K{ openmp_tls | openmp_notls }

These options specify whether or not to apply automaticparallelization.

-K{parallel | noparallel}

These options specify whether or not to avoid calculation error ofa real type or a complex type operation that is caused by thedifference of the parallel number of threads.

-K{ parallel_fp_precision |parallel_nofp_precision }

The -Kparallel_iteration option specifies that only the loops whichare analyzed to iterate N times or more are targets of parallelization.

-Kparallel_iteration=N

1 <= N <= 2147483647

The -Kparallel_strong option specifies to parallelize all loopswhich can be parallelized automatically without estimatingparallelization effects.

-Kparallel_strong

These options specify whether or not to apply reductionoptimization.

-K{reduction | noreduction}

- 11 -


These options specify whether or not to make the parallel regionextended.

-K{region_extension | noregion_extension}

These options specify whether or not to create the thread-safeobject program.

-K{threadsafe | nothreadsafe}

The -Kvisimpact option specifies to create object programsoptimized for VISIMPACT (Virtual Single Processor byIntegrated Multicore Parallel Architecture).

-Kvisimpact

Table 2.9 [Options for debug and tools]


These options specify whether or not to output information fordebug into object programs.

{ -g | -g0 }

These options specify to debug a serial program. -H{a | e | f | o | s | u | x}

The -Ncheck_global option specifies to check followingcharacteristics in program units during compilation.

- The procedure characteristics between external proceduredefinitions and references.

- The procedure characteristics between external proceduredefinitions and interface body.

- Size of common blocks.

-Ncheck_global

The -Ncheck_intrfunc option specifies to perform intrinsicfunction error processing.

-Ncheck_intrfunc

These options specify whether or not to use the hook functioncalled from a specified location.

-N{ hook_func | nohook_func }

These options specify whether or not to use the hook functioncalled at regular time interval.

-N{ hook_time | nohook_time }

These options specify whether or not to output the statementnumber where the user defined procedure which causes an error iscalled and the statement number where an error occurred in theprocedure.

-N{line | noline}

These options specify to debug Fortran source code.

-Eg

-Nquickdbg [=dbg_arg]

dbg_arg: {{ argchk | noargchk }| { subchk |nosubchk }| { undef | undefnan | noundef } |{ inf_detail | inf_simple }}

These options specify whether or not to output the runtimeinformation.

-N{ rt_tune | rt_notune }

-Nrt_tune_func

-Nrt_tune_loop[=kind]

kind: {all|innermost}

These options specify whether or not to trap floating-pointexceptions.

-N{Rtrap | Rnotrap}

Table 2.10 [Options for messages]


The -f option specifies the lowest error level of diagnosticmessages.

-f{i | w | s | msg_num}

- 12 -


These options specify to check whether a source code conforms toa specific Fortran standard.

-v{03d | 03e | 03o | 03s}

The -Ec option specifies to output diagnostic messages whenexpressions may not be evaluated accurately.

-Ec

These options specify whether or not to output information ofapplied optimizations.

-K{optmsg[=level] | nooptmsg}

level: {1 | 2}

These options specify whether to cancel the compilation if thecompiler forecasts that it takes a long time to compile the program.

-N{ cancel_overtime_compilation |nocancel_overtime_compilation }

The -Ncheck_cache_arraysize option specifies to inspect the sizesof arrays during the compilation and issue a level i diagnosticmessage if the sizes could be a cause an impact to executionperformance.

-Ncheck_cache_arraysize

These options specify whether or not to output the name of file andprogram unit which are currently processed.

-N{compdisp | nocompdisp}

The -Nmaxserious option specifies to stop compilation whenserious errors are detected more than maxnum times.

-Nmaxserious=maxnum

These options specify to output information of compilation into .lstfile.

-Nlst[={a | d | i | m | p | t | x}]

-Nlst_out=file

-Q[{a | d | i | m | ofile | p | t | x}]

Table 2.11 [Options for linker]


The -shared option specifies to create shared objects. -shared

The -l option specifies to link libname.so or libname.a -lname

These options specify whether or not to use largepage function. -K{largepage | nolargepage}

The -L option specifies to search libraries from directory. -Ldirectory

Table 2.12 [Other options]


The -c option specifies to suppress calling a linker. -c

These options specify to change method to rename externalprocedures.

-ml={cdecl|frt}

-mldefault={cdecl | frt}

The -mlcmain option specifies to select C main program name. -mlcmain={main | MAIN__}

The -o option specifies to name output file exe_file. -oexe_file

The -I option specifies to search include files and moduleinformation (.mod) files from directory.

-Idirectory

The -M option specifies to output module information (.mod) filesto directory, or to search .mod files from directory.

-Mdirectory

The -P option specifies to perform only preprocessing. -P

The -S option specifies to output assembly source. -S

The -V option specifies to output the version information of eachtool.

-V

The -W option specifies to use arg1, arg2, ... as options for tool. -Wtool,arg1[,arg2]...

- 13 -


These options specify to output absolute path of tools and optionsused by frtpx command.

-#

-###

Option List

[ -c ] [ -fmsg_lvl ] [ { -g | -g0 } ] [ -lname ] [ -ml=target ] [ -mlcmain=main_program ] [ -mldefault=target ][ -oexe_file ] [ -shared ] [ -vmsg_arg ] [ -xinl_arg ][ -Aobj_arg ] [ -Ctyp_arg ] [ -Dname[=tokens] ] [ -E ] [ -Emsg_arg ] [ -Fixed ] [ -Free ] [-Fwide ][ -Hdbg_arg ] [ -Idirectory ] [ -Kopt_arg ] [ -Ldirectory ] [ -Mdirectory ][ -Nsrc_arg ] [ -O[opt_lvl] ] [ -P ] [ -Q[lst_arg] ] [ -S ] [ -SSL2 ] [ -SSL2BLAMP ][ -Uname ] [ -V ] [ -Wtool,arg1[,arg2]... ] [ -Xlan_lvl ] [ -# ] [ -### ]

-c

The -c (compile only) option suppresses calling the linking phase. If the -c option is specified, object programs are created, but anexecutable program is not created. It is ignored if the -P and -S option is specified. If only object files are specified by file, it ismeaningless to invoke the compile command with the -c option.

-f msg_lvl msg_lvl: { i | w | s | msg_num }

The -f option specifies the lowest error level of diagnostic messages output for Fortran compiler errors.

Each diagnostic message has a prefix of the form jwdxxxxz-y message, where xxxx is the message serial number, z is the messagekind, and y is the message error level. The y can be either i (information message), w (warning message), s (serious error), or u(unrecoverable error). The -f option is the default.

See Section "4.1 Compilation Output" for information about diagnostic messages output during compilation.

If -fmsg_num is specified, the particular error messages whose message number is msg_lvl are suppressed.

The msg_num argument can coexist with other arguments (including other msg_num arguments). To specify msg_num and otherarguments for the -f option, separate the arguments by a comma (,).

Example:

$ frtpx -fw,1040,2005 a.f95

i

The -fi option specifies to output all levels of diagnostic messages.

w

The -fw option specifies to output diagnostic messages only for error levels w, s, and u.

s

The -fs option specifies to output diagnostic messages only for error levels s and u.

msg_num

The -fmsg_num suppresses particular information and warning diagnostic messages specified by the msg_num argument. Themsg_num argument is the four-digit message number; more than one msg_num may be specified. The msg_num argument cancoexist with any other arguments, but is ignored when -fs is specified.

{ -g | -g0 }

These options direct whether to generate the debugging information used by the debugger.

-g

When this option is effective, the debugging information is generated in object programs.

The user can perform source-level debugging of programs compiled with this option.

When this option is effective, and the option which makes the -O1 or higher option effective is not specified, the -O0 option iseffective.

When this option is effective, the -Karray_merge_common[=name], -Karray_merge_local, -Karray_subscript, and -Klto optionsare ignored.

- 14 -

Note the following restrictions when using debugger with program compiled with -O1 or higher option:

- Debugging of the variables is guaranteed at the entry of each procedure, and debugging of only the dummy arguments, variablesdeclared in modules, and variables declared in common blocks is guaranteed. And note the following restrictions:

- Debugging of the dummy arguments of which type is quadruple precision REAL, single precision COMPLEX, doubleprecision COMPLEX, quadruple precision COMPLEX, derived type, or polymorphic is restricted.

- Debugging of the dummy arguments of which type is CHARACTER whose length type parameter is not constant, arraywhose lower or upper bound is not constant, or assumed-shape array is restricted.

- Debugging of the dummy arguments which are not referred in its procedure is restricted.

- Debugging of the local variables is not guaranteed.

- Debugging of the inlined procedures is restricted.

- The displaying of line number is not guaranteed.

Compile the program with the compiler option -O0 when using debugger without the restrictions above.

See "Debugger User's Guide" for the debugger specification.

-g0

This option invalidates the -g option.

-lname

The -l option specifies to search the library libname.so or libname.a. The position of this option within the command line is important.This is because libraries are searched for in the order in which the other libraries and object files appear within the command line. Thisoption and its argument are passed to the linker.

-ml=target target: { cdecl | frt }

The -ml option specifies to change how to process external procedure name. cdecl or frt can be specified for target. See Section "11.3Rules for Processing Fortran Procedure Names" for details.

-mlcmain=main_program main_program: { main | MAIN__ }

The -mlcmain option specifies C main program name. C main program name is a function name specified for main_program. main orMAIN__ can be specified for the function name. See Section "11.5.1 Passing Control First to a C Program" for details.

-mldefault=target target: { cdecl | frt }

The -mldefault option specifies to change how to process default external procedure names according to target. cdecl or frt can bespecified for target. See Section "11.3 Rules for Processing Fortran Procedure Names" for details.

-o exe_file

The -o option specifies to name the output file exe_file. Unless either the -c or -S option is specified, the output file is the executableprogram. If the -c option is specified, the output file is the object program. If the -S option is specified, the output file is the assemblyprogram.

-shared

The -shared option specifies to create a shared library. This option is passed for linker program.

-v msg_arg msg_arg: { 03d | 03e | 03o | 03s }

The -v option specifies to diagnose whether a source code conforms to a specific Fortran standard.

To specify plural arguments for the -v option, separate the arguments by a comma (,).

03d

The -v03d option specifies to issue diagnostic messages if the source code contains any deleted features in Fortran 2003.

03e

The -v03e option specifies to issue diagnostic messages if the source code contains any Fortran 2003 features not also in Fortran95.

- 15 -

03o

The -v03o option specifies to issue diagnostic messages if the source code contains any obsolescent features in Fortran 2003.

03s

The -v03s option specifies to issue diagnostic messages if the source code contains any nonstandard features, other than decrementalfeatures in Fortran 2003.

-xinl_arg inl_arg: { - | proc_name[,proc_name]... | stmt_no | dat_szK | dir=dir_name[,dir=dir_name]... | 0 }

The -x option specifies to apply inline expansion to user-defined procedures. Inline-expansion improves the execution performanceas -K series options. Either of the -O1 option or higher must be specified together. To specify plural arguments for the -x option,separate the arguments by a comma.

The -x0 option is default.

This optimization increases the compilation time and memory requirements. The more optimizations are applied, the larger the objectprogram size is. Specify the -Koptmsg=2 option to know whether or not this optimization is applied.

-

The -x- option specifies to choose the targets of inline expansion automatically from the user defined functions.

proc_name[,proc_name]...

The -xproc_name[,proc_name]... option specifies to apply inline expansion only to user-defined procedures which is specified byproc_name arguments. Specifying only the names of frequently called procedures can reduce compilation time and memoryrequirements significantly. To specify proc_name and other arguments (including other proc_name arguments) for the -x option,separate the arguments by a comma. If stmt_no or dat_szK is specified in addition to proc_name, the compiler applies inlineexpansion for procedures which meet either condition.

To specify module procedure name as proc_name, the module name and a period must be specified immediately before withoutintervening blanks.

Similarly, to specify internal procedure name as proc_name, the host procedure name and a period must be specified immediatelybefore without intervening blanks.

Moreover, if the host procedure is module procedure, the module name and a period must be specified immediately before the hostname without intervening blanks.

Example: Specifying internal procedure name

$ frtpx -xsub.insub a.f90

stmt_no

The -xstmt_no option specifies to apply inline expansion for user-defined procedures having stmt_no or fewer executablestatements. The argument stmt_no must be an integer value between 1 and 2147483647.

dat_szK

The -xdat_szK option specifies to apply inline expansion only to user-defined procedures whose all local arrays are up to dat_szK.The argument dat_sz must be an integer value between 1 to 2147483. The letter K, represents kilobytes (not kibibytes), must beadded after the value of dat_sz. The default dat_sz is unlimited.

The -xdat_szK option applies inline expansion for user-defined procedures whose local array size is less than the specifieddat_szK value.

If inline expansion for user-defined procedures is applied, the generated object program includes the data referenced in the user-defined procedure. If this data includes a huge local array, the generated object program can be quite large. To control this, specifythe -x dat_szK option.

dir=dir_name[,dir=dir_name]...

The-xdir=dir_name option specifies to apply inline expansion for procedures defined in the file under the directory specified bythe argument dir_name or referenced in the file currently compiled. However, these files are necessary to be the followingconditions:

- These files whose suffix is .f, .for, .f90, .f95, or .f03. And,

- 16 -

- These files must not need to specify -Cpp option.

When the -xdir=dir_name option is specified, files under the dir_name directory are compiled temporarily by the compiler optionswhich are enabled during the compilation of input file, and these are target of inline expansion.

To specify dir=dir_name and other arguments (including other dir=dir_name arguments) for the -x option, separate the argumentsby a comma.

In that case, the other sub options are effective on the all files under the specified directory.

Specifying the -xdir=dir_name option, needs plenty time and large work area for compilation.

0

The -x0 option specifies not to apply inline expansion.

-A obj_arg obj_arg: { A | E | e | T | U | d | i | p | q | w | y | z }

The arguments of the -A option are A, E, e, T, U, d, i, m, p, q, w, y and z. These arguments can be specified at the same time.

A

The -AA option specifies not to align data and derived type in common blocks. See Section "5.2 Correct Data Boundaries" fordetails.

E

The -AE option specifies not to treat special characters such as backslash (\) character as control-characters. The -AE option isdefault.

e

The -Ae option specifies to treat special characters such as backslash (\) as control-characters.

T

The -AT option does not apply implicit typing and specifies the null mapping for all the letters. This does not affect any IMPLICITstatements.

U

The -AU option indicates that names are to be interpreted in a case-sensitive manner. Be careful of the following:

- This causes the program to be interpreted in a nonstandard manner. In ISO standard Fortran, letters are not case sensitive.

- References to intrinsic module entities, service subroutines, service functions or the external procedure names of the multi-operation functions must be in lowercase letters.

- The spelling of all declarations and references to an intrinsic procedure must be the same.

- In the IMPLICIT statement, the lowercase letters specified in a letter specifier list are equivalent to the corresponding uppercaseletters.

- When you debug using the debugger, user-defined procedure names are case sensitive and variable names are not case sensitive.

d

The -Ad option improves constants, variables, and functions (intrinsic, statement, external, module and internal functions) of defaultreal and default complex types to double precision. See Section "5.3.1 Precision Improving" for information about this function.

The -CcR8R16, -CcR4R8, -CcdRR8, -CcdDR16, -Ccd4d8, and -Cca4a8 options cannot be used with the -Ad option.

i

The -Ai option changes the default for integer type data from four-byte to two-byte. Two-byte integer type data will be used for:

- Integer constants with absolute values from 0 to 32767

- Data typed using the default implicit typing rules and having symbolic names beginning with I, J, K, L, M, or N

- Data without kind selector declared in an INTEGER statement

- Data resulting from intrinsic functions INT (without one-byte integer, four-byte integer, and eight-type integer argument),NINT, IDNINT, ICHAR, MAX1, MIN1, LEN, and INDEX

- 17 -

The -CcI4I8, -CcdII8, -Ccd4d8, and -Cca4a8 options cannot be used with the -Ai option.

p

The -Ap option specifies a default of private accessibility. It is same as explicitly specifies PRIVATE statement.

q

The -Aq option improves constants, variables, and functions (intrinsic, statement, external, module and internal functions) ofdouble-precision real and double-precision complex types to quadruple precision. See Section "5.3.1 Precision Improving" forinformation about this function.

The -CcR8R16, -CcR4R8, -CcdRR8, -CcdDR16, -Ccd4d8, and -Cca4a8 options cannot be specified with this option.

w

The -Aw option causes the IACHAR, ACHAR, IBITS, and ISHFTC to be intrinsic procedures even if the -X6 or -X7 option is ineffect.

y

The -Ay option improves the kind (precision) of an unsigned real literal constant on the right side of an assignment statement tothe kind of the variable on the left side. For example,

REAL(8) AA=1.23456789012345

The above program is equivalent to the following program if the -Ay option is specified.

REAL(8) AA=1.23456789012345_8

z

The -Az option appends the null character (\0) to the end of each character argument passed to an external procedure. This allowsa character string to be passed correctly to a C function, such as strlen. The length of the character string argument does not includethe null character. For example, the result of following program is 1234 4 , even if the -Az option is specified.

CALL SUB('1234')ENDSUBROUTINE SUB(ARG)CHARACTER(LEN=*) ARGPRINT *,ARG , LEN(ARG)END

-C typ_arg typ_arg: { l | cpp | fpp | pp | cdII8 | cI4I8 | cdLL8 | cL4L8 | cdRR8 | cR4R8 | cd4d8 | ca4a8 | cdDR16 | cR8R16 }

The -C option controls numerical accuracy and data alignment, the C preprocessor, the Fortran preprocessor, and the evaluation ofdata types. To specify more than one argument for the -C option, separate the arguments by a comma (,).

l 0 <= l <= 15

The l arguments of the -C option determine the effect of rounding errors on the results of floating-point data operations. The l mustbe a number between 0 to15. When the -Cl option is specified, the right most l bits of the mantissa are set to zero each time a valueis assigned to a default real, double-precision real, quadruple-precision real, default complex, double-precision complex, orquadruple-precision complex type variable by an assignment statement. (For complex types, these bits are set to zero in both thereal and imaginary parts.) See Section "5.3.2 Precision Lowering and Analysis of Errors" for information about this function.

cpp

When it meets either of the following requirements, the -Ccpp option calls the C language preprocessor.

- The suffix of Fortran source file is F, FOR, F90, F95, or .F03.

- The -Cpp option is specified.

The C language preprocessor processes Fortran source as follows.

- The names are to be interpreted in a case-sensitive manner.

- 18 -

- The preprocessor ignores the comment for Fortran, the continuing line for Fortran, and the column number in fixed sourceform for Fortran.

If you want to know more detail, see "Appendix E Preprocessor".

fpp

When it meets either of the following requirements, the -Cfpp option calls the Fortran preprocessor.

- The suffix of Fortran source file is F, FOR, F90, F95, or .F03.

- The -Cpp option is specified.

The Fortran source is processed as follows.

- The names are to be interpreted in a case-sensitive manner.

- The comment for Fortran is given to priority more than the comment for C language.

- The preprocessor recognizes the comment for Fortran, the continuing line for Fortran, and the column number in fixed sourceform for Fortran.

If you want to know more detail, see "Appendix E Preprocessor".

pp

The -Cpp option calls the C language or Fortran preprocessor. The preprocessor of default is Fortran. If you want to use the Cpreprocessor, specify -Cpp and -Ccpp options.

cdII8

The -CcdII8 option interprets the default integer variables, constants, and functions as eight-byte integers.

If the -CcdII8 option is specified, INT, IFIX, IDINT, IQINT, NINT, IDNINT, and IQNINT specific intrinsic functions must notbe used as an actual argument.

If the -CcdII8 option is specified, the specific intrinsic functions NINT and IDNINT must not be used as right side expression ofpointer assignment statement and component of the structure constructor.

The -Ai option cannot be used with this option. For example,

INTEGER ::AINTEGER(4) ::BA=2B=2_4

is evaluated as follows if the -CcdII8 option is specified.

INTEGER(8) ::AINTEGER(4) ::BA=2_8B=2_4

cI4I8

The -CcI4I8 option interprets any four-byte integer type (whether default four-byte or explicitly declared four-byte) as eight-byteinteger type. It applies to variables, constants, and functions.

If the -CcI4I8 option is specified, INT, IFIX, IDINT, IQINT, NINT, IDNINT, and IQNINT specific intrinsic functions must notbe used as an actual argument.

If the -CcI4I8 option is specified, the specific intrinsic functions NINT and IDNINT must not be used as right side expression ofpointer assignment statement and component of the structure constructor.

The -Ai option cannot be used with this option. For example,

INTEGER ::AINTEGER(4) ::BA=2B=2_4

- 19 -

is evaluated as follows if the -CcI4I8 option is specified.

INTEGER(8) ::AINTEGER(8) ::BA=2_8B=2_8

cdLL8

The -CcdLL8 option interprets default logical variables, constants, and functions as eight-byte logicals.

If the -CcdLL8 option is specified, the specific intrinsic function BTEST must not be used as an actual argument.

For example,

LOGICAL ::ALOGICAL(4) ::BA=.TRUE.B=.TRUE._4

is evaluated as follows if the -CcdLL8 option is specified.

LOGICAL(8) ::ALOGICAL(4) ::BA=.TRUE._8B=.TRUE._4

cL4L8

The -CcL4L8 option interprets any four-byte logical type (whether default four-byte or explicitly declared four-byte) as an eight-byte logical type. It applies to variables, constants, and functions.

If the -CcL4L8 option is specified, the specific intrinsic function BTEST must not be used as an actual argument. For example,

LOGICAL ::ALOGICAL(4) ::BA=.TRUE.B=.TRUE._4

is evaluated as follows if the -CcL4L8 option is specified.

LOGICAL(8) ::ALOGICAL(8) ::BA=.TRUE._8B=.TRUE._8

cdRR8

The -CcdRR8 option interprets the default real type as double-precision real type and interprets the default complex type as double-precision complex type. It applies to variables, constants, and functions.

If the -CcdRR8 option is specified, the specific intrinsic functions REAL, FLOAT, SNGL, and SNGLQ must not be used as anactual argument.

The -Ad and -Aq option cannot be used with this option. For example,

REAL ::AREAL(4) ::BCOMPLEX ::CCOMPLEX(4) ::DA=2.0B=2.0_4C=(2.0,2.0)D=(2.0_4,2.0_4)

is evaluated as follows if the -CcdRR8 option is specified.

- 20 -

REAL(8) ::AREAL(4) ::BCOMPLEX(8) ::CCOMPLEX(4) ::DA=2.0_8B=2.0_4C=(2.0_8,2.0_8)D=(2.0_4,2.0_4)

cR4R8

The -CcR4R8 option is equivalent to specifying the -Ad option. It interprets any default real or complex type (whether default four-byte or explicitly declared four byte) as double-precision real or complex type. It applies to variables, constants, and functions.

If the -CcR4R8 option is specified, the specific intrinsic functions REAL, FLOAT, SNGL, and SNGLQ must not be used as anactual argument.

The -Ad and -Aq option cannot be used with this option. For example,

REAL ::AREAL(4) ::BCOMPLEX ::CCOMPLEX(4) ::DA=2.0B=2.0_4C=(2.0,2.0)D=(2.0_4,2.0_4)

is evaluated as follows if the -CcR4R8 option is specified.

REAL(8) ::AREAL(8) ::BCOMPLEX(8) ::CCOMPLEX(8) ::DA=2.0_8B=2.0_8C=(2.0_8,2.0_8)D=(2.0_8,2.0_8)

cd4d8

The -Ccd4d8 option is equivalent to specifying all of the -CcdII8, -CcdLL8, and -CcdRR8 options. For example,

INTEGER ::AINTEGER(4) ::BLOGICAL ::CLOGICAL(4) ::DREAL ::EREAL(4) ::FCOMPLEX ::GCOMPLEX(4) ::HA=2B=2_4C=.TRUE.D=.TRUE._4E=2.0F=2.0_4G=(2.0,2.0)H=(2.0_4,2.0_4)

is evaluated as follows if the -Ccd4d8 option is specified.

INTEGER(8) ::AINTEGER(4) ::B

- 21 -

LOGICAL(8) ::CLOGICAL(4) ::DREAL(8) ::EREAL(4) ::FCOMPLEX(8) ::GCOMPLEX(4) ::HA=2_8B=2_4C=.TRUE._8D=.TRUE._4E=2.0_8F=2.0_4G=(2.0_8,2.0_8)H=(2.0_4,2.0_4)

ca4a8

The -Cca4a8 option is equivalent to specifying all of the -CcI4I8, -CcL4L8, and -CcR4R8 options. For example,

INTEGER ::AINTEGER(4) ::BLOGICAL ::CLOGICAL(4) ::DREAL ::EREAL(4) ::FCOMPLEX ::GCOMPLEX(4) ::HA=2B=2_4C=.TRUE.D=.TRUE._4E=2.0F=2.0_4G=(2.0,2.0)H=(2.0_4,2.0_4)

is evaluated as follows if the -Cca4a8 option is specified.

INTEGER(8) ::AINTEGER(8) ::BLOGICAL(8) ::CLOGICAL(8) ::DREAL(8) ::EREAL(8) ::FCOMPLEX(8) ::GCOMPLEX(8) ::HA=2_8B=2_8C=.TRUE._8D=.TRUE._8E=2.0_8F=2.0_8G=(2.0_8,2.0_8)H=(2.0_8,2.0_8)

cdDR16

The -CcdDR16 option interprets the double-precision real type as quadruple-precision real type. It applies to variables, constants,and functions.

If the -CcdDR16 option is specified, the specific intrinsic functions DFLOAT, DBLE, DBLEQ, DREAL, and DPROD must notbe used as an actual argument.

If the -CcdDR16 option is specified, the specific intrinsic function DPROD must not be used as right side expression of pointerassignment statement and component of the structure constructor.

- 22 -

The -Ad, and -Aq option cannot be used with this option. For example,

DOUBLE PRECISION ::AREAL(8) ::BA=2.0D0B=2.0_8

is evaluated as follows if the -CcdDR16 option is specified.

REAL(16) ::AREAL(8) ::BA=2.0_16B=2.0_8

cR8R16

The -CcR8R16 option is equivalent to specifying the -Aq option. It interprets any double-precision real type as quadruple-precisiontype and interprets any double-precision complex type as quadruple-precision complex type. It applies to variables, constants, andfunctions.

If the -CcR8R16 option is specified, the specific intrinsic functions DFLOAT, DBLE, DBLEQ, DREAL, and DPROD must notbe used as an actual argument.

If the -CcR8R16 option is specified, the specific intrinsic function DPROD must not be used as right side expression of pointerassignment statement and component of the structure constructor.

The -Ad, and -Aq option cannot be used with this option. For example,

DOUBLE PRECISION ::AREAL(8) ::BCOMPLEX(8) ::CA=2.0D0B=2.0_8C=(2.0_8,2.0_8)

is evaluated as follows if the -CcR8R16 option is specified.

REAL(16) ::AREAL(16) ::BCOMPLEX(16) ::CA=2.0_16B=2.0_16C=(2.0_16,2.0_16)

-D name[=tokens]

The -D option specifies to define name as #define directives do. If tokens is omitted, name is defined to be 1.

If both -Dname and -Uname are specified, the -Uname option invalidate -Dname option regardless of the order. Even if -D is notspecified, the following macros are defined by default.

Macro Value Note

__FUJITSU 1

__arch64__ 1

__unix 1

__sparc 1

__sparcv9 1

__linux 1

unix 1

sparc 1

- 23 -

Macro Value Note

linux 1

__frt_version 700

_OPENMP 201107 This macro is defined when the -Kopenmp option is effective

-E

The -E option specifies to perform only preprocessing to the specified source file.

The result of preprocessing is output to the standard-output.

The -Cpp option must be specified together unless a suffix of the source file is .F, .FOR, .F90, .F95 or .F03.

The -P option cannot be specified together.

-E msg_arg msg_arg: { c | g }

The -E option tells the Fortran compiler to output diagnostic messages about operations for which round off may affect the results.

The arguments of the -E option are c and g. These arguments may be combined, e.g., -Ecg.

c

The -Ec option specifies to output diagnostic messages when expressions may not be evaluated accurately. During compilation,this option also outputs level i diagnostic messages for arithmetic IF statements with real expressions.

See Section "4.1.1 Compilation Diagnostic Messages" for information about diagnostic messages output during compilation. SeeSection "6.3.2 Real Data Operations" for information about relational operations.

g

This function increases the compilation time. When the -Eg option is specified, the -Ncheck_global, -Haefosux and -O0 optionsare in effect.

-Fixed

The -Fixed option specifies to compile a Fortran source program as a fixed-form source.

-Free

The -Free option specifies to compile a Fortran source program as a free-form source.

-Fwide

The -Fwide option specifies to compile a Fortran source program as a fixed-form source whose length of all lines is 255.

-H dbg_arg dbg_arg: { a | e | f | o | s | u | x }

The -H option specifies the level of run-time error checking. During compilation some errors can be detected in Fortran object programsand error messages are generated. Program execution automatically detects some errors, and the arguments a, e, f, o, s, u and x can bespecified to broaden the range of errors checked. Arguments a, e, f, o, s, u and x can be combined. These arguments can be specifiedat the same time as -Haefosu.

The -Eg option is supported to check the program error, too. It checks characteristics of a procedure between procedure definition andreference, between procedure definition and interface body, and size of common block during compilation.

If the -H option is specified, the default optimization is -O0. The -O option may be specified after the -H option to make the -O optioneffective. See Section "8.2 Debugging Functions" for information about the debug function.

a

The -Ha option compares the numbers, types, attributes, and sizes of actual and dummy arguments. The types of function resultsare also checked during execution. If a mismatch occurs during execution, the corresponding diagnostic message is output. SeeSection "8.2.1.1 Checking Argument Validity (ARGCHK)" for information about the debug function.

e

The -He options checks shape conformance when an array assignment statement is executed. If a mismatch of shape conformanceoccurs during execution, the corresponding diagnostic message is output. See Section "8.2.1.4 Shape Conformance Check(SHAPECHK)" for information about the debug function.

- 24 -

f

The -Hf option checks whether the file is connected with two devices or more at the same time in the input/output statement, andwhether the input/output statement is called in addition is inspected by the function while executing the input/output statement.

If a mismatch occurs during execution, the corresponding diagnostic message is output. See Section "8.2.1.7 Simultaneous OPENand I/O Recursive Call CHECK (I/O OPEN)" for information about the debug function.

o

The -Ho option checks the follows:

- Two dummy arguments are overlapped and the part of overlap is changed.

- When an assumed-size array with INTENT(OUT) attribute is referenced during execution, checks to see if the variable isdefined.

- When a variable with SAVE attribute is referenced during execution, checks to see if the variable is defined.

If a mismatch occurs during execution, the corresponding diagnostic message is output. For information about the debug function,see Section "8.2.1.6 Overlapping Dummy Arguments and Extend Undefined CHECK (OVERLAP)".

When the -Ho option is specified, -Ha option and -Hu option are in effect. See Section "8.2.1.1 Checking Argument Validity(ARGCHK)" for information about the -Ha debug function. See Section "8.2.1.3 Checking References for Undefined Data Items(UNDEF)" for information about the -Hu debug function.

s

When an array section, array element, or substring is referenced during execution, the -Hs option checks the value of each subscriptor substring expression to see if it is within the declared range. The same check is done during compilation when possible. If thevalue is not within the declared range, a diagnostic message is output. See Section "8.2.1.2 Checking Subscript and SubstringValues (SUBCHK)" for information about the debug function.

u

When a variable outside a common block and a module is referenced during execution, the -Hu option checks to see if the variableis defined, If the referenced variable is undefined, a diagnostic message is output. See Section "8.2.1.3 Checking References forUndefined Data Items (UNDEF)" for information about the u argument.

x

The -Hx option checks to see if the variable declared in a module or in a common block is defined when the variable is referencedduring execution, to see if the pointer is associated when a pointer is referenced during execution, and to see if the incrementationparameter of the DO construct, stride of the FORALL, incrementation parameter of array constructor implied do control and strideof subscript-triplet is not 0.

If the -Hx option is specified, the -Hu option is in effect.

Be careful to the following when the -Hx option using:

- All program units that have a definition or initialization for common block object shall be compiled with -Hx option.

- All program units that uses the module shall be compiled with -Hx option, if the module is compiled with -Hx option.

See Section "8.2.1.5 Extended Checking of UNDEF (EXTCHK)" for information about the x argument.

-I directory

The -I option specifies directory names to be searched for files specified in an INCLUDE line or for module information (.mod) files.

More than one -I option may be specified. If multiple -I options are specified, the path names are searched in the order specified.

INCLUDE file retrieval is performed in the following order:

1. The directory of the files containing the INCLUDE lines

2. The directory in which frtpx is being executed

3. Directories specified as the argument of an -I option

The compilation information of a module is searched in the following order:

1. Directories specified as the argument of a -M option

- 25 -

2. The directory in which frtpx is being executed

3. Directories specified as the argument of an -I option

-Kopt_arg opt_arg: { adr44 | adr64 | noalias[=spec] | arraypad_const[=N] | arraypad_expr=N | array_merge |array_merge_common[=name] | array_merge_local | array_merge_local_size=N | { array_private | noarray_private } |array_subscript | array_subscript_element=N | array_subscript_elementlast=N | array_subscript_rank=N |array_subscript_variable='ary_nm(rank,rank[,rank]...)' | array_transform | { auto | noauto } | { autoobjstack | noautoobjstack }| { calleralloc[=level] | nocalleralloc } | commonpad[=N] | dalign | nodalign } | { dynamic_iteration | nodynamic_iteration } | { eval| noeval } | fast | { fed | nofed } | { fenv_access | nofenv_access } | { FLTLD | NOFLTLD } | { fp_contract | nofp_contract } |{ fp_relaxed | nofp_relaxed } | { fsimple | nofsimple } | { HPC_ACE | HPC_ACE2 } | { ilfunc | noilfunc } | independent=proc_name| instance=N | { intentopt| nointentopt } | { largepage | nolargepage } | { loop_blocking[=N]} | loop_noblocking } | { loop_fission| loop_nofission } | { loop_fission_if | loop_nofission_if } | { loop_fusion | loop_nofusion } | { loop_interchange |loop_nointerchange } | { loop_part_parallel | loop_nopart_parallel } | { loop_part_simd | loop_nopart_simd } | { loop_versioning| loop_noversioning } | { lto | nolto } | { mfunc[=level] | nomfunc } | { nf | nonf } | { ns | nons } | { ocl | noocl } | { omitfp | noomitfp }| { openmp | noopenmp } | {openmp_assume_norecurrence | openmp_noassume_norecurrence } |{ openmp_ordered_reduction | openmp_noordered_reduction } | { openmp_tls | openmp_notls } | options | { optmsg[=level] |nooptmsg } | { parallel | noparallel } | { parallel_fp_precision | parallel_nofp_precision } | parallel_iteration=N | parallel_strong| { pic | PIC } | { preex | nopreex } | noprefetch | prefetch_cache_level=N | { prefetch_conditional prefetch_noconditional } |{ prefetch_indirect | prefetch_noindirect } | { prefetch_infer | refetch_noinfer } | prefetch_iteration=N |prefetch_iteration_L2=N | prefetch_line=N | prefetch_line_L2=N | { prefetch_sequential[=kind] | prefetch_nosequential } |{ prefetch_sequential | prefetch_nosequential } | { prefetch_stride | refetch_nostride } | { prefetch_strong | prefetch_nostrong }| { prefetch_strong_L2 | prefetch_nostrong_L2 } | { reduction | noreduction } | { region_extension | noregion_extension } |{ shortloop=N | noshortloop } | { simd[=level] | nosimd } | { simd_separate_stride | simd_noseparate_stride } | { static_fjlib |nostatic_fjlib } | { striping[=N] | nostriping } | { swp | noswp } | swp_strong | { temparraystack | notemparraystack } | { threadsafe| nothreadsafe } | { unroll[=N] | nounroll } | { uxsimd | nouxsimd } | visimpact | vppocl | { XFILL[=N] | NOXFILL } }

The -K options controls various aspects of optimization in order to improve execution performance of object programs. It is specifiedusing arguments, each controlling a different area of optimization. Multiple arguments can be specified, separated with commas.

adr44

The -Kadr44 option specifies to create object files in which absolute address is expressed as 44-bit data.

adr64

The -Kadr64 option specified to create object files in which absolute address is expressed as 64-bit data.

noalias[=spec]

The -Knoalias directs the compiler to perform optimization under the assumption that different pointers or target variables in thesource code do not assign values to the same memory area. 's' can be specified in spec.

When 's' is specified for spec, the compiler performs optimization only on variables with pointer attributes that meet Fortranstandards. When -Knoalias is specified, the compiler performs optimization and automatic parallelization under the aboveconditions, promoting optimization and automatic parallelization of pointer variables. Note that the optimizations may not bepromoted when the association status of the pointer variables is changed in the loop.

If this option is specified, and the pointers in the source program do not meet the above conditions, the executable program mayoutput an incorrect result.

arraypad_const[=N] 1 <= N <= 2147483647

If the first dimension of an array has explicit bounds and the bounds expression is constant, -Karraypad_const directs to executepadding to align N elements. When N is not specified, the compiler will determine the padding volume for each array. Padding isthe act of creating spaces within arrays.

This cannot be specified at the same time as the -Karraypad_expr=N.

See Section "9.12 Effects of Applying Optimization to Change Shape of Array" for points to note when specifying this option.

arraypad_expr=N 1 <= N <= 2147483647

Regardless of whether the bounds expression is a constant, -Karraypad_expr directs to executes padding to align N elements inarrays where the first dimension has explicit bounds.

This cannot be specified at the same time as the -Karraypad_const[=N] option.

- 26 -

See Section "9.12 Effects of Applying Optimization to Change Shape of Array" for points to note when specifying this option.

array_merge

The -Karray_merge option is equivalent to specifying the -Karray_merge_local and -Karray_merge_common options.

See Section "9.15 Merging Array Variables" for points to note when specifying this option.

array_merge_common[=name]

The -Karray_merge_common option directs that multiple arrays within common blocks are to be merged. If the -Ncheck_global,-Nquickdbg=argchk,-Nquickdbg=undef, -Nquickdbg=undefnan, -Nquickdbg=subchk, -g or -H option is set, -Karray_merge_common[=name] option will be ignored. Specify the common block name in the name.

If this name is omitted, the operation will target all arrays within named common blocks.

array_merge_local

The -Karray_merge_local option specifies to whether multiple local arrays are to be merged. If the -Ncheck_global, -Nquickdbg=argchk,-Nquickdbg=undef, -Nquickdbg=undefnan, -Nquickdbg=subchk, -g or -H option is set, the -Karray_merge_local option will be invalid. -Karray_merge_local_size=1000000 will also take effect at the same time.

array_merge_local_size=N 2 <= N <= 2147483647

The -Karray_merge_local_size option directs that the size of the local arrays to be merged are N bytes or greater. The -Karray_merge_local option must be specified together.

{ array_private | noarray_private }

These options direct whether or not to localize arrays in loops. This option is effective when -Kparallel option is effective. Thedefault is -Knoarray_private.

array_private

The -Karray_private option directs to localize arrays in loops. The -Karray_private option is effective when -Kparallel optionis effective.

noarray_private

The -Knoarray_private option directs not to localize arrays in loops.

array_subscript

The -Karray_subscript option causes an array dimension shift to be performed on:

- Allocatable arrays with 4 or more dimensions.

- Arrays with 4 or more dimensions that have 10 or fewer elements in the last dimension but more than 100 elements in the otherdimensions.

See Section "9.14 The Dimension Shift of the Array Declaration" for points to note when specifying this option. If the -Ncheck_global, -Nquickdbg=argchk,-Nquickdbg=undef, -Nquickdbg=undefnan, -Nquickdbg=subchk, -g or -H option is set, the -Karray_subscript option will be invalid.

-Karray_subscript_element=100, -Karray_subscript_elementlast=10, and -Karray_subscript_rank=4 will take effect at the sametime.

array_subscript_element=N 2 <= N <= 2147483647

The -Karray_subscript_element option directs that the number of elements other than the last dimension of the array (for whichdimension shift is to be performed) is N or more. -Karray_subscript_element is effective only when -Karray_subscript is set. Thisoption is not effective in allocatable arrays.

array_subscript_elementlast=N 2 <= N <= 2147483647

The -Karray_subscript_elementlast option directs that the number of elements of the last dimension of the array (for whichdimension shift is to be performed) is N or fewer. -Karray_subscript_elementlast is effective only when the -Karray_subscriptoption is set. This option is not effective in allocatable arrays.

array_subscript_rank=N 2 <= N <= 30

The -Karray_subscript_rank option directs that the number of elements of the array (for which dimension is to be performed) is Nor more. This is effective only when the -Karray_subscript option is set.

- 27 -

array_subscript_variable='ary_nm(rank,rank[,rank]...)' 1 <= rank <= 30

The -Karray_subscript_variable option specifies to perform dimension shift on the array specified by 'ary_nm(rank,rank[,rank]...)'.The argument ary_nm is the name of array to optimize, and the argument "rank" is the position of the dimension to be shifted. Thenumber of specified dimension positions must be the same as the dimension number of the argument ary_nm. The same positionof the dimension cannot be specified two times or more. This is effective only when the -Karray_subscript option is set, and willinvalidate the values specified in the -Karray_subscript_element=N, -Karray_subscript_elementlast=N, and -Karray_subscript_rank=N options.

array_transform

The -Karray_transform option is equivalent to specifying the -Karraypad_const, -Karray_merge, and -Karray_subscript.

{ auto | noauto }

These options direct whether to treat local variables (other than those with the SAVE attribute or initial values) as automaticvariables. The automatic variables become undefined when the procedure ends. The default is -Knoauto.

auto

Local variables, other than those with the SAVE attribute or initial values, are treated as automatic variables and allocated tothe stack. See Section "9.11 Effects of Allocating on Stack" for points to note when specifying this option.

noauto

Local variables, other than those with the SAVE attribute or initial values, are not treated as automatic variables.

{ autoobjstack | noautoobjstack }

These options specify whether to allocate automatic data on the stack or the heap. The default is -Knoautoobjstack.

autoobjstack

The -Kautoobjstack option directs that automatic data objects are allocated to the stack. See Section "9.11 Effects of Allocatingon Stack" for points to note when specifying this option.

noautoobjstack

The -Knoautoobjstack option directs that automatic data objects are allocated to the heap.

{ calleralloc[=level] | nocalleralloc } level: { 1 | 2 }

These options direct whether the function result area is allocated by the caller or called, when referencing user-defined arrayfunction. The default is -Knocalleralloc.

Specify 1 or 2 for level.The default value is 1. Function results matching the conditions are allocated by the caller according tolevel. With function results that do not meet the condition, the area is allocated dynamically during execution of the function, andthe caller deallocates the area after return.

Specifying this option may reduce overhead at execution caused by the dynamic allocation and deallocation of the function resultarea.

The same option must be specified for all files in which the array function is defined or referenced. If this option is specifieddifferently or different levels are specified across the caller and called, the program may not work correctly.

calleralloc=1

When a user-defined array function is referenced, the caller of the function allocates the area for function results that meet thefollowing conditions:

- The result variables of the function are primitive data types.

- The result variables of the function are explicit-shape arrays and all explicit bounds declarations are constant expressions.

- When the result variables of the function are character types, the character length parameter of those variables are constantexpressions.

calleralloc=2

In addition to the conditions of calleralloc=1, the result area that meets the following conditions will be allocated by the callerof the function.

- -Karraypad_expr=N is not valid.

- 28 -

- The result variables of the function are primitive data types.

- When the result variables of the function are character types, the character length parameter of those variables are constantexpressions.

- At the declaration of the function, all the explicit bounds declarations are constant expressions or intrinsic function SIZEreferences.

- The argument of the intrinsic function SIZE is a derived array name without an OPTIONAL attribute.

- The intrinsic function SIZE does not have a second argument.

nocalleralloc

The -Knocalleralloc option directs to allocate the area of the function result dynamically whenever the function is executedwith user-defined array function references. It also deallocates the area at the caller after return.

commonpad[=N] 4 <= N <= 2147483644

The -Kcommonpad option specifies to pad array in common blocks or declaration part of module to use cache efficiently.

If the source file which includes the common block or module is compiled with the -Kcommonpad option, other source files whichinclude the same common block or using module must be compiled with same -Kcommonpad. N is specified by bytes.

When N is omitted, the compiler determines automatically suitable value. When the common block or declaration part of modulecontain only array variables, the pads are effective.

{ dalign | nodalign }

These options direct whether to generate instructions assuming that eight-byte integer data, double-precision real data, quadruple-precision real data, double-precision complex data, or quadruple-precision complex data referred to by common blocks, dummyarguments, or pointers are aligned on eight-byte boundaries. The default is -Knodalign.

dalign

Instructions are generated assuming that the object program being generated is aligned on eight-byte boundaries. If -Kdalignis used, all the procedures comprising the executable program must be compiled with -Kdalign specified.

nodalign

Instructions are generated assuming that the object program being generated is not aligned on eight-byte boundaries.

{ dynamic_iteration | nodynamic_iteration }

These options direct whether to dynamically select a loop to be parallelized considering the number of threads when both the innerloop and the outer loop are candidates in a nested loop. This option is effective only when the -Kparallel option is set. The defaultis -Knodynamic_iteration.

dynamic_iteration

Loops are dynamically selected to be parallelized considering the number of threads when both the inner loop and the outerloop are candidates in a nested loop. The nest level for the loops is up to three.

nodynamic_iteration

Loops are not dynamically selected to be parallelized when both the inner loop and the outer loop are candidates in a nestedloop.

{ eval | noeval }

These options direct whether to perform optimization that modifies how operations are evaluated for object programs. Note thatperforming this optimization may cause side effects in the execution result. The numerical operations do not conform to IEEE 754.The default is -Knoeval.

eval

The -Keval option modifies how operations are evaluated for object programs. When -Keval is set, -Kfsimple is valid. Inaddition, when this option and -Kparallel are set, -Kfsimple and -Kreduction are valid.

This option is effective only when the -O1 option or higher is set.

- 29 -

noeval

The -Knoeval option does not modify how operations are evaluated for object programs. When -Knoeval is set, -Knofsimpleand -Knoreduction option are valid.

fast

The -Kfast option is for creating object programs to be executed at the highest speed on the target machine.

This is equivalent to specifying the -O3 -Kdalign,eval,fp_contract,fp_relaxed,ilfunc,mfunc,ns,omitfp,prefetch_conditional options.As -Kfast includes the -Keval, -Kfp_contract, -Kfp_relaxed, -Kilfunc, and -Kmfunc options, there may be side effects on thecalculation results. The numerical operations do not conform to IEEE 754.

{ fed | nofed }

These options direct whether to create an executable program that uses floating-point exception disable mode.

These options must be set at linking. The default is -Knofed.

These options are valid only when the -KHPC_ACE2 option is set.

fed

An executable program that uses floating-point exception disable mode is created.

The program can be executed at high speed owing to disabling the floating-point exception.

The numerical operations do not conform to IEEE 754 and programs that handle floating-point exceptions run incorrectly. Itis necessary to use this option carefully enough because all functions and libraries that are called from the executable programoperate in floating-point exception disable mode. This option includes the function of the -Kns option regardless of whether tospecify the -Kns or -Knons option.

The -Kfed option is valid only when the -O1 option or higher is set. If both the -Kfed option and the -NRtrap option or the -Nquickdbg=undefnan option are specified, the one specified last is valid.

nofed

An executable program that does not use floating-point exception disable mode is created.

{ fenv_access | nofenv_access }

These options direct whether or not the program may access floating point rounding mode and also direct whether or not therounding mode is the default. Some floating point optimization may be suppressed when this option is valid.

The default is -Knofenv_access.

fenv_access

The -Kfenv_access specifies that the program can access floating point rounding mode or the rounding mode is not the default.Some floating point optimization may be suppressed.

nofenv_access

The -Knofenv_access specifies that the program does not access the floating point rounding mode and this directs that therounding mode is the default.

{ FLTLD | NOFLTLD }

These options direct whether to perform optimization using non-faulting load instructions on object programs. The non-faultingload instruction is one of the load instructions which do not cause an exception whether or not the given address is correct. Usingthis instruction extends the range of optimization and will promote further optimization. The default is -KFLTLD.

These options are valid only when the -KHPC_ACE option is set.

FLTLD

The -KFLTLD option specifies not to perform optimizations using non-faulting-load instructions.

NOFLTLD

The -KNOFLTLD option specifies to perform optimizations using non-faulting-load instructions. -KNOFLTLD is effectiveonly when the -O1 option or higher is set.

- 30 -

{ fp_contract | nofp_contract }

These options direct whether or not to perform optimization using "Floating-Point Multiply-Add/Subtract" floating point operationinstructions on object programs.

Rounding errors may occur when these instructions are generated. The default is -Knofp_contract.

fp_contract

Optimization using Floating-Point Multiply-Add/Subtract floating point operation instructions is performed on object programs.


nofp_contract

Optimization using Floating-Point Multiply-Add/Subtract floating point operation instructions is not performed on objectprograms.

{ fp_relaxed | nofp_relaxed }

These options specify whether or not to perform optimizations using reciprocal approximation operation instructions on single-precision or double precision floating point divisions or SQRT functions. This optimization may introduce rounding errors,compared to when the normal division instructions or SQRT instructions are used. The numerical operations do not conform toIEEE 754. The default is -Knofp_relaxed.

fp_relaxed

The -Kfp_relaxed option directs that reciprocal approximation operation instructions and Floating-Point Multiply-Add/Subtractfloating point operation instructions are used on single-precision or double precision floating point division or SQRT functions.-Kfp_relaxed is effective only when the -O1 option or higher is set.

nofp_relaxed

The -Knofp_relaxed option directs that normal division instructions or SQRT instructions are used for single-precision or doubleprecision floating point division or SQRT functions.

{ fsimple | nofsimple }

These options direct whether to perform simplification of floating point operation on object programs. For example, operationssuch as "x*0" will be simplified to "0".The default is -Knofsimple. Using this optimization the numerical operations do not conformto IEEE 754.

fsimple

Simplification of floating point operation on object programs is performed. -Kfsimple is effective only when the -O1 option orhigher is set.

nofsimple

Simplification of floating point operation on object programs is not performed.

{ HPC_ACE | HPC_ACE2 }

These options specify whether to generate the object file using instruction set for either HPC-ACE or HPC-ACE2. The -KHPC_ACE2 option is default.

HPC_ACE

-KHPC_ACE option specifies to generate the object file using HPC-ACE instruction set.

If this option is specified, note the following:

- The object file generated by this option cannot operate on FX10 system.

- If this option is specified to the main program, the dynamic library linked at runtime is the library for HPC-ACE. So, theexecution performance may decrease.

- The SIMD instruction for HPC-ACE2 executes 4 operations at the same time. On the other hand, the SIMD instruction forHPC-ACE executes 2 operations at the same time, so the execution performance of the loop which SIMD optimization is appliedto may decrease by this option.

HPC_ACE2

-KHPC_ACE2 option specifies to generate the object file using HPC-ACE2 instruction set.

- 31 -

{ ilfunc | noilfunc }

These options direct whether to perform inline expansion for intrinsic functions and operations. The compiler determines whetherto perform inline expansion, based on this option, the data type, the code size, etc. Note that performing this optimization maycause rounding errors in the result. Further, if the argument of the intrinsic function corresponds to the "Cause of error" describedin Section "8.1.5 Intrinsic Function Error Processing", error messages are not output and the execution results are not guaranteed.The default is -Knoilfunc. Using this optimization the numerical operations do not conform to IEEE 754.

The following shows the intrinsic functions and operations for which inline expansion can be performed.

- Intrinsic functions of the single-precision and double-precision real type: ATAN, ATAN2, COS, EXP, EXP10, LOG, LOG10,SIN, and TAN

- Exponentiations of the single-precision and double-precision real type (Both the base and exponent must be the same type)

- Intrinsic function of the single-precision complex type: EXP

- Intrinsic functions of the double-precision complex type: ABS and EXP

ilfunc

Inline expansion is performed for intrinsic functions and operations. -Kilfunc is effective only when the -O1 option or higheris set. In addition, reciprocal approximation operation and Floating-Point Multiply-Add/Subtract floating point operation maybe used, even if -Knofp_contract and -Knofp_relaxed option are set.

noilfunc

Inline expansion is not performed for intrinsic functions and operations.

independent=pgm_nm

The -Kindependent option directs that the procedure reference specified in pgm_nm always performs as a sequential process, evenif the procedure reference is specified within the parallelized DO-loop, and would be subject to automatic parallelization. -Kindependent is effective only when the -Kparallel option is set.

Procedure pgm_nm must be compiled with the -Kthreadsafe option specified. If the -Kthreadsafe option is not specified, the resultsof the execution cannot be guaranteed.

To specify module procedure name as pgm_nm, the module name and a period must be specified immediately before withoutintervening blanks.

Similarly, to specify internal procedure name as pgm_nm, the host procedure name and a period must be specified immediatelybefore without intervening blanks.

For an internal procedure-name that is within module procedures, the module-name must also be specified.

Example: To specify the internal procedure-name "insub" that is within the parent procedure sub

$ frtpx -Kparallel,independent=sub.insub a.f90

This option shares points to note with the optimization indicator INDEPENDENT. See Section "12.2.3.3.2 Nested ParallelProcessing" and Section "12.2.3.3.4 Notes on Using Optimization Control Line" for details.

instance=N 2 <= N <= 512

The -Kinstance=N option specifies the number of threads at execution. Specify a value between 2 and 512 for N. -Kinstance iseffective only when the -Kparallel option is set. If the value of N differs from the number of threads at execution, execution willbe aborted. See Section "12.2.3.3.1 Caution when specifying compile-time option -Kparallel,instance=N" for details.

{ intentopt | nointentopt }

These options direct whether the optimization which uses the intent is performed in the program unit including procedure call,which procedure call includes the actual argument associated with the dummy argument to which INTENT attribute is specified.

When the -O0 option is set, the default is -Knointentopt. When the -O1 option or higher is set, the default is -Kintentopt.

intentopt

The optimization which uses the intent is performed in the program unit including procedure call, which procedure call includesthe actual argument associated with the dummy argument to which INTENT attribute is specified.


- 32 -

nointentopt

The optimization which uses the intent is not performed in the program unit including procedure call, which procedure callincludes the actual argument associated with the dummy argument to which INTENT attribute is specified.

{ largepage | nolargepage }

These options specifies whether or not to create executable program where use large-page function. These options must be set atlinking.

See Section "D.5 Large-Page Function" for information about large-page function. The -Klargepage option is default.

largepage

-Klargepage option specifies to create the executable program which the large-page function is applied.

nolargepage

-Knolargepage option specifies to create the executable program which the large-page function isn't applied.

{ loop_blocking[=N] | loop_noblocking } 2 <= N <= 10000

These options direct whether to perform the loop blocking optimization. Specify the block size between 2 and 10000 for N. If avalue was not specified for N, the compiler will automatically determine a suitable value. The default is -Kloop_blocking, whenthe -O2 option or higher is set.

loop_blocking[=N]

Loop blocking is performed. -Kloop_blocking is valid only when the -O2 option or higher is set.

loop_noblocking

Loop blocking is not performed.

{ loop_fission | loop_nofission }

These options direct whether to perform loop fission:

- The optimization which divides a loop into plural loops.

The default is -Kloop_fission, when the -O2 option or higher is set.

loop_fission

Loop fission is performed. -Kloop_fission is valid only when the -O2 option or higher is set.

loop_nofission

Loop fission is not performed.

{ loop_fission_if | loop_nofission_if }

These options direct whether to perform loop fission optimization as much as possible before and after IF constructs in loops. The-Kloop_fission_if option is valid when the -Kloop_fission option is set. The -Kloop_nofission_if is default.

See Section "9.1.2.7 Loop Fission before and after IF constructs in loops(LFI)" for details.

loop_fission_if

The -Kloop_fission_if option specifies to perform the loop fission optimization as much as possible before and after IF constructsin loops.

loop_nofission_if

The -Kloop_nofission_if option specifies not to perform the loop fission optimization as much as possible before and after IFconstructs in loops, but to perform normal loop fission.

{ loop_fusion | loop_nofusion }

These options direct whether to perform loop fusion, which combines adjacent loops. The default when the -O2 option or higheris set is -Kloop_fusion.

loop_fusion

Loop fusion is performed. -Kloop_fusion is valid only when the -O2 option or higher is set.

- 33 -

loop_nofusion

Loop fusion is not performed.

{ loop_interchange | loop_nointerchange }

These options specify whether or not to exchange loops.

When either the -O2 option or higher is specified, -Kloop_interchange option is effective and derived.

When the -Kloop_nointerchange option is specified, -Kloop_interchange option is ignored.

{ loop_part_parallel | loop_nopart_parallel }

When a loop contains statements to which parallelization can be applied and statements to which parallelization cannot be applied,the loop is divided into two or more loops and automatic parallelization is applied to one of the loops. This optimization is appliedto the innermost loop. Note that compile time and execution time may increase. The -Kloop_part_parallel option is effective onlywhen both the -Kloop_fission option and the -Kparallel option are set. The default is -Kloop_nopart_parallel.

loop_part_parallel

The automatic parallelization which needs dividing loops is performed.

loop_nopart_parallel

The automatic parallelization which needs dividing loops is not performed.

{ loop_part_simd | loop_nopart_simd }

When a loop contains instructions to which SIMD extensions can be applied and instructions to which SIMD extensions cannotbe applied, the loop is divided into two or more loops and SIMD extensions are applied to one of the loops. This optimization isapplied to the innermost loop. Note that compile time and execution time may increase. The -Kloop_part_simd option is effectiveonly when both the -Kloop_fission option and the -Ksimd option are set. The default is -Kloop_nopart_simd option.

loop_part_simd

The optimization using SIMD extensions which needs dividing loops is performed.

loop_nopart_simd

The optimization using SIMD extensions which needs dividing loops is not performed.

{ loop_versioning | loop_noversioning }

These options specify whether to perform optimization by the loop versioning.

The loop versioning may promote optimizations such as SIMD, software pipelining, or automatic parallelization.

Note that the size of the object program and compile time may increase because the loop versioning generates two loops. Moreover,the execution performance of the program may decrease due to the overhead of judgement for the choice of generated loops.

These options are effective only when the -O2 option or higher is set. The default is -Kloop_noversioning.

For details about the loop versioning, see Section "9.1.2.8 Loop Versioning".

loop_versioning

The -Kloop_versioning option specifies to perform optimization by the loop versioning.

loop_noversioning

The -Kloop_noversioning option specifies not to perform optimization by the loop versioning.

{ lto | nolto }

These options direct whether to perform link time optimization. The default is -Knolto.

If the -g, -S or -Wa option is set, -Klto option will be invalid.

If -Klto option and -xdir=dir_name option are specified, -xdir=dir_name option will be invalid.

See Section "9.1.2.10 Link Time Optimization" for this function.

lto

Link time optimization is performed. -Klto option is effective only when the -O1 option or higher is set.

- 34 -

nolto

Link time optimization is not performed.

{ mfunc[=level] | nomfunc } level: { 1 | 2 | 3 }

These options direct whether to perform optimization by converting intrinsic functions and operations to multi-operation functions."Multi-operation functions" are optimized intrinsic functions which perform same operations for multiple arguments at a singlecalling to improve performance. However, they may introduce potential rounding errors due to a difference in algorithms. Thisoption is effective only when the optimization level -O2 option or higher is set. Specify 1, 2, or 3 for level. The default value forlevel is 1. The default is -Knomfunc.

The following shows the intrinsic functions and operations that can be converted.

Function/Operation

nameSingle

precision realtype

Doubleprecision real

type

Singleprecision

complex type

Doubleprecision

complex type

Integer type

ACOS o (*1) o (*1) o (*2) o (*2) -

ACOSH x x o (*2) o (*2) -

ASIN o (*1) o (*1) o (*2) o (*2) -

ASINH x x o (*2) o (*2) -

ATAN o o o (*2) o (*2) -

ATAN2 o o - - -

ATANH x x o (*2) o (*2) -

COS o o o (*2) o (*2) -

COSH o (*2) o (*2) o (*2) o (*2) -

COSQ x x o (*2) o (*2) -

ERF o (*1) o (*1) - - -

ERFC o (*1) o (*1) - - -

EXP o o o (*2) o -

EXP10 o o - - -

LOG o o o (*2) o (*2) -

LOG10 o o - - -

SIN o o o (*2) o (*2) -

SINH o (*2) o (*2) o (*2) o (*2) -

SINQ x x o (*2) o (*2) -

TAN x x o (*2) o (*2) -

TANH o (*2) o (*2) o (*2) o (*2) -

exponentiation o o o (*2) o (*2) o (*2)

ISHFT - - - - o (*2)

o: To be targetx: Not to be target-: No Intrinsic function

*1) These multi-operation functions execute sequential instructions internally. The execution performance improves becauseoptimizations, such as branch optimization, are applied to the internal instructions. The improvement is inferior to that of thenormal multi-operation functions.

- 35 -

*2) These multi-operation functions call the corresponding intrinsic functions four times internally. Although the performanceof Multi-operation functions is the same as when calling the intrinsic functions four times, SIMD optimization to loops includingintrinsic functions is promoted by using Multi-operation functions.

mfunc=1

The -Kmfunc=1 option specifies to use multi-operation functions with 4 multiplicities.

mfunc=2

In addition to the functions used by "-Kmfunc=1", this option specifies to use multi-operation functions with 8 or moremultiplicities. This function requires a large stack area.

mfunc=3

In addition to "-Kmfunc=2", this option specifies to also use the multi-operation functions with 8 or more multiplicities forloops including IF constructs. Note that the performance will be worse than when "-Kmfunc=1" or "-Kmfunc=2" is specified,when the TRUE rate of the IF construct is low. Note that performing this optimization may cause side effects in the executionresult. See Section "9.16.2 Effects of Compiler Option -Kmfunc=3" for information on the side effects. This function requiresa large stack area.

nomfunc

Optimization by converting intrinsic functions and operations to multi-operation functions is not performed.

{ nf | nonf }

These options direct whether to perform optimization using the Non-Faulting mode on object programs. Using the Non-Faultingmode extends the range of speculative execution optimization, and optimization can be promoted.

Even if the given address is not correct, the load instruction in the Non-Faulting mode does not cause an exception.

When -Ksimd={2|auto} option and -Knf option are specified at the same time, the load instructions that are not speculative executionin the loop either might operate in the Non-Faulting mode. The default is -Knonf.

These options are valid only when the -KHPC_ACE2 option is set.

nf

Optimization using the Non-Faulting mode is done to the object program.

-Knf option is valid only when the -O1 option or higher is set.

nonf

Optimization using the Non-Faulting mode is not done to the object program.

{ ns | nons }

These options direct whether to initialize the FPU with non-standard floating-point mode. These options must be set at linking.The default is -Knons. When the -Kfed option is effective, the function of the -Kns option always becomes effective.

ns

The FPU is initialized with non-standard floating-point mode. When the FPU is initialized with non-standard floating-pointmode, underflow value may become zero to get more performance. Moreover, the numerical operations do not conform to IEEE754. -Kns is effective only when the -O1 option or higher is set.

nons

The FPU is not initialized with non-standard floating-point mode.

{ ocl | noocl }

These options direct whether to enable optimization control lines. See Section "9.10 Using Optimization control line (OCL)" forinformation on optimization control lines. The default is -Knoocl.

ocl

Optimization control lines are enabled. -Kocl is effective only when the -O1 option or higher is set.

noocl

Optimization control lines are disabled.

- 36 -

{ omitfp | noomitfp }

These options specify whether or not to keep the frame pointer in a register from procedure calling. The -Knoomitfp is default.

omitfp

When the -Komitfp option is effective, the frame pointer in a register is not kept. When this option is set, the trace backinformation is not kept. Either of the -O1 option or higher must be specified together.

noomitfp

When the -Knoomitfp option is effective, the frame pointer in a register is kept.

{ openmp | noopenmp }

These options direct whether to enable OpenMP directives. See Section "12.3 Parallelization by OpenMP Specifications" forinformation on this function. The default is -Knoopenmp.

openmp

The OpenMP directives are enabled. When the -Kopenmp option is set, the -Kthreadsafe and -Kauto options will also be set.However, the local variable of the main program will not be allocated to the stack even if the -Kauto option is set.

The -Knoauto and -Knothreadsafe options are valid if they are specified after the -Kopenmp option. Caution should be takenwhen the -Knoauto and -Knothreadsafe options are specified at the same time as the -Kopenmp option. (See Section "12.3Parallelization by OpenMP Specifications").

Even when only object programs are specified as file names (that is, when only linking is to be performed), -Kopenmp mustbe specified if an object program that was compiled by with the -Kopenmp option specified is included.

noopenmp

OpenMP directives are treated as comments.

{ openmp_assume_norecurrence | openmp_noassume_norecurrence }

These options direct whether or not to promote optimizations by assuming that array elements of the chunk size have no datadependency over iteration for the loop with the OpenMP DO directive.

When -Kopenmp_assume_norecurrence is effective, SIMD extensions, software pipelining, and other optimizations are promoted.

These options are applied to the innermost loops with the OpenMP DO directive.

These options are effective only when the -Kopenmp option and -O2 option or higher is set. The default is -Kopenmp_noassume_norecurrence.

openmp_assume_norecurrence

This option directs to assume that array elements of the chunk size have no data dependency over iteration to promoteoptimizations.

openmp_noassume_norecurrence

This option directs not to assume that array elements of the chunk size have no data dependency over iteration.

{ openmp_ordered_reduction | openmp_noordered_reduction }

These options specify whether or not to fix the operation order of the reduction operations same as in the numerical order of thethreads at the end of the region for which the reduction clause of OpenMP was specified. If the number of threads used is identical,by fixing the order of the reduction operations same as in the numerical order of the threads, identical results will be always obtained.However, rounding errors may occur when the operation order is changed by the effect of the scheduling of a loop construct witha dynamic or a guided schedule kind, or sections construct.

The -Kopenmp_noordered_reduction option is default.

openmp_ordered_reduction

The -Kopenmp_ordered_reduction option specifies to fix the operation order of the reduction operations same as in the numericalorder of the threads at the end of the region for which the reduction clause was specified. Note that execution performance maydecrease compared to when the -Kopenmp_ordered_reduction is invalidated. This option is valid only when the -Kopenmpoption is in effect.

- 37 -

openmp_noordered_reduction

The -Kopenmp_noordered_reduction option specifies not to fix the operation order of the reduction operation at the end of theregion for which the reduction clause was specified.

{ openmp_tls | openmp_tls }

These options specify whether or not to allocate the threadprivate variable on the Thread-Local Storage. The default is -Kopenmp_notls.

If the threadprivate variable is passed between Fortran and C/C++ programs, it is necessary to specify the -Kopenmp_tls option.

If the -Kopenmp_tls option is used, all the source programs comprising the executable program must be compiled with the -Kopenmp_tls option.

openmp_tls

The -Kopenmp_tls option directs that threadprivate variables are allocated to the Thread-Local Storage. This option is effectiveonly when the -Kopenmp option is set.

openmp_notls

The -Kopenmp_notls option directs that threadprivate variables are allocated to the heap.

options

The -Koptions directs that lines starting with "!options" are treated as compiler directives. See Section "2.5 Compiler DirectiveLines" for information on compiler directives.

{ optmsg[=level] | nooptmsg } level: { 1 | 2 }

These options direct whether to output messages about optimization status. Specify 1 or 2 for level. The default value is 1. Thedefault is -Koptmsg=1.

optmsg=1

Messages are output indicating that the optimization performed may cause side effects in the execution results.

optmsg=2

Along with the messages from optmsg=1, messages are output indicating that optimization functions for automaticparallelization, SIMD optimization, and loop unrolling have performed.

nooptmsg

Output of optimization status messages is suppressed.

{ parallel | noparallel }

These options direct to performed automatic parallelization. However, if the effect of parallel execution is not expected, automaticparallelization is not performed. This option induces the -O2 and -Kregion_extension option. If the -O3 option is set, however, theoptimization level will be 3. If the -H or -Eg option is set, the -Kparallel option will be ignored. The default is -Knoparallel.

See Section "12.2 Automatic Parallelization" for information on this option.

Even when only object programs are specified as file names (that is, when only linking is to be performed), this option must bespecified if an object program that was compiled with the -Kparallel option specified is included.

parallel

The -Kparallel option directs that automatic parallelization is performed.

noparallel

The -Knoparallel option directs that automatic parallelization is not performed.

{ parallel_fp_precision | parallel_nofp_precision }

These options specify whether to avoid calculation error of a real type or a complex type operation that is caused by the differenceof the parallel number of threads. The default is -Kparallel_nofp_precision.

parallel_fp_precision

The -Kparallel_fp_precision option specifies to avoid calculation error of a real type or a complex type operation that is causedby the difference of the parallel number of threads.

- 38 -

This option is effective when -Kparallel or -Kopenmp option is effective.

When this option and -Kopenmp option are set, -Kopenmp_ordered_reduction option is valid.

When this option is set, the execution performance may decrease because the part of optimization is restricted.

Note that the calculation error of a real type or a complex type operation that is caused by the difference of the parallel numberof threads even if this option is valid when the reduction clause of OpenMP is specified.

parallel_nofp_precision

The -Kparallel_nofp_precision option specifies not to avoid calculation error of a real type or a complex type operation that iscaused by the difference of the parallel number of threads.

parallel_iteration=N 1 <= N <= 2147483647

The -Kparallel_iteration option directs that only loops determined at compilation to have a number of iterations exceeding N willbe subject to parallelization. -Kparallel_iteration is effective only when the -Kparallel option is set.

Example :

REAL(KIND=4),DIMENSION(10000,5) :: A,B,CDO J=1,5 DO I=2,10000 A(I,J) = B(I,J) + C(I,J) END DOEND DO

When -Kparallel_iteration=6 option is set, the outer loop will not be subject to parallelization because its iteration is 5.

parallel_strong

The -Kparallel_strong option directs to paralellize all loops for which it is detected that paralellization can be performed withoutestimating the effects of the paralellization. This option includes the -Keval and -Kpreex options.

Aside from this point, the functions and notes are the same as for the -Kparallel option.

{ pic | PIC }

These options direct to generate position-independent code (PIC).

If the -KPIC option is specified, a slower object with a longer instruction sequence is generated, however the number of uniqueexternal symbols that can be referenced when linking increases. If the -Kpic option is specified, a faster object with a shorterinstruction sequence is generated, but the number of unique external symbols that can be referenced when linking decreases. Inthese cases, the number of unique external symbols is the total of all libraries that are referenced simultaneously.

If both -KPIC and -Kpic are specified, the one specified last is valid. This option is valid only when specified during compilation.

{ preex | nopreex }

These options direct whether to evaluate invariant expressions in advance. Note that this optimization may cause errors, asinstructions that may not have been executed based on the logic of the program are likely to be executed. To identify whether ornot this optimization was performed, see the diagnostic messages at compilation.

The default is -Knopreex.

preex

Invariant expressions are evaluated in advance. -Kpreex is effective only when the -O1 option or higher is set.

nopreex

Invariant expressions are not evaluated in advance.

noprefetch

The -Knoprefetch option directs that objects using prefetch instructions are not created.

prefetch_cache_level=N N: { 1 | 2 | all }

The -Kprefetch_cache_level option directs which cache level is to have data prefetched. -Kprefetch_cache_level is effective onlywhen the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect option is set. Specify 1, 2, or all for N. The default is -Kprefetch_cache_level=all.

- 39 -

prefetch_cache_level=1

The -Kprefetch_cache_level=1 option directs to prefetch data in the first level cache. Normal prefetch instructions are used.

prefetch_cache_level=2

The -Kprefetch_cache_level=2 option directs to prefetch data only using the second level cache.

prefetch_cache_level=all

Functionality of prefetch_cache_level=1 and prefetch_cache_level=2 are valid simultaneously. Faster prefetch can be achievedby using the two types of prefetch instruction in combination.

{ prefetch_conditional | prefetch_noconditional }

These options specify whether or not to generate prefetch instructions for array data used in blocks in IF constructs or CASEconstructs. This option requires that the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect option is set. The defaultis -Kprefetch_noconditional.

prefetch_conditional

Prefetch instructions are generated for array data used in blocks in IF constructs and CASE constructs.

prefetch_noconditional

Prefetch instructions are not generated for array data used in blocks in IF constructs and CASE constructs.

{ prefetch_indirect | prefetch_noindirect }

These options direct whether to create an object that uses prefetch instructions for array data that accessed indirectly (list access)within a loop. This option is effective only when the -O1 option or higher is set. The default is -Kprefetch_noindirect. Specifyingthis option may cause non-faulting load instructions to be created.

prefetch_indirect

Prefetch instructions are generated for array data accessed indirectly (list access) within a loop.

prefetch_noindirect

Prefetch instructions are not generated for array data accessed indirectly (list access) within a loop.

{ prefetch_infer | prefetch_noinfer }

These options direct whether to create prefetch instructions even if the prefetching distance is unknown. This option requires thatthe -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect option is set. The default is -Kprefetch_noinfer.

prefetch_infer

Prefetch instructions for sequential access are created, even if the prefetching distance is unknown.

prefetch_noinfer

Prefetch instructions for sequential access are not created if the prefetching distance is unknown.

prefetch_iteration=N 1 <= N <= 10000

The -Kprefetch_iteration option specifies to prefetch data that is referenced or defined after N iterations of the loop.

As this option is only for prefetch instructions that prefetch to the first level cache, it is valid only when any option of the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect is set, and, the -Kprefetch_cache_level=1 or the -Kprefetch_cache_level=all option is set.

This cannot be specified with the -Kprefetch_line=N option.

prefetch_iteration_L2=N 1 <= N <= 10000

The pretech_iteration_L2 option specifies to prefetch data that is referenced or defined after N iterations of the loop.

As this option is only for prefetch instructions that prefetch to the second level cache, it is valid only when any option of the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect is set, and, the -Kprefetch_cache_level=2 or the -Kprefetch_cache_level=all option is set.

This cannot be specified at the same time as the -Kprefetch_line_L2=N option.

- 40 -

prefetch_line=N 1 <= N <= 100

The -Kprefetch_line option specifies to prefetch data which are referenced or defined after N cache line(s).

As this option is only for prefetch instructions that prefetch to the first level cache, it is valid only when the -Kprefetch_sequentialor -Kprefetch_indirect is set, and, the -Kprefetch_cache_level=1 or the -Kprefetch_cache_level=all option is set.

This cannot be specified at the same time as the -Kprefetch_iteration=N option.

prefetch_line_L2=N 1 <= N <= 100

The -Kprefetch_line_L2 option specifies to prefetch data that is referenced to or defined in a line after N lines.

As this option is only for prefetch instructions that prefetch to the second level cache, it is valid only when the -Kprefetch_sequentialor -Kprefetch_indirect option is valid, and, the -Kprefetch_cache_level=2 or the -Kprefetch_cache_level=all option is set.

This cannot be specified at the same time as the -Kprefetch_iteration_L2=N option.

{ prefetch_sequential[=kind] | prefetch_nosequential } kind: {auto|soft}

These options direct whether to generate prefetch instructions for array data accessed sequentially within a loop. Specify auto orsoft for kind. The default value of kind is auto. The default when the -O0 or -O1 option is set is -Kprefetch_nosequential. Thedefault when the -O2 option or higher is set is -Kprefetch_sequential.

prefetch_sequential=auto

The compiler automatically selects whether to use hardware-prefetch or to create prefetch instructions for array data that isaccessed sequentially within a loop. -Kprefetch_sequential=auto is effective only when the -O1 option or higher is set.

prefetch_sequential=soft

The compiler does not use hardware-prefetch, but rather creates prefetch instructions for array data that is accessed sequentiallywithin a loop. -Kprefetch_sequential=soft is effective only when the -O1 option or higher is set.

prefetch_nosequential

Prefetch instructions are not generated for array data that is accessed sequentially within a loop.

{ prefetch_stride | prefetch_nostride }

These options direct whether to generate prefetch instructions for array data that is accessed with a stride larger than the cache linesize used in the loop. This option is effective only when the -O1 option or higher is set. The default is -Kprefetch_nostride.

prefetch_stride

Prefetch instructions are generated for array data that is accessed with a stride larger than the cache line size used in a loop.This includes loops with prefetch addresses not defined at compilation.

prefetch_nostride

Prefetch instructions are not generated for array data that is accessed with a stride larger than the cache line size used in a loop.

{ prefetch_strong | prefetch_nostrong }

These options direct whether to create strong prefetch instructions for the first level cache.

As this option is only for prefetch instructions that prefetch to the first level cache, it is valid only when any option of the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect is set, and, the -Kprefetch_cache_level=1 or the -Kprefetch_cache_level=all option is set.

The default is -Kprefetch_nostrong.

prefetch_strong

Prefetch instructions created for the first dimension cache are strong prefetch.

prefetch_nostrong

Prefetch instructions created for the first dimension cache are not strong prefetch.

{ prefetch_strong_L2 | prefetch_nostrong_L2 }

These options direct whether to create strong prefetch instructions for the second level cache.

- 41 -

As this option is only for prefetch instructions that prefetch to the second level cache, it is valid only when any option of the -Kprefetch_sequential, -Kprefetch_stride, or -Kprefetch_indirect is set, and, the -Kprefetch_cache_level=2 or the -Kprefetch_cache_level=all option is set.

The default is -Kprefetch_strong_L2.

prefetch_strong_L2

Prefetch instructions created for the second level cache are strong prefetch.

prefetch_nostrong_L2

Prefetch instructions created for the second level cache are not strong prefetch.

{ reduction | noreduction }

These options direct whether to perform the reduction optimization. This option is effective only when the -Kparallel option is set.The default is -Knoreduction. Using this optimization the numerical operations do not conform to IEEE 754. See Section "12.2Automatic Parallelization" and Section "12.2.3.1.8 Loop Reduction" for information about this function.

reduction

Reduction optimization is performed.

noreduction

Reduction optimization is not performed.

{ region_extension | noregion_extension }

These options specify whether or not to expand parallel-region to reduce parallelization overhead. This option is effective onlywhen the -Kparallel option is set. See Section "12.2.3.1.11 Extension of Parallel Region" for information on this function. Thedefault is -Knoregion_extension.

region_extension

Parallelization overhead is reduced by expanding parallelization.

noregion_extension

Parallelization is not expanded.

{ shortloop=N | noshortloop } 2 <= N <= 10

These options direct whether some optimizations for short loops are applied to all innermost loops in the program.

-Kshortloop=N specifies to modify some optimization functions at innermost loops assuming its iteration is few as the specifiedN. N can be specified from 2 to 10. The default is -Knoshortloop. -Kshortloop=N is valid only when the -O2 option or higher isset.

shortloop=N

-Kshortloop=N directs that some optimizations for short loops are applied to all innermost loops. -Kshortloop=N is valid onlywhen the -O2 option or higher is set.

At the innermost loop with N iterations, -Kshortloop=N performs the followings:

- controlling the number of loop unrolling

- suppression of software pipelining

- suppression of loop blocking

- suppression of loop striping

- suppression of automatic parallelization

- suppression of XFILL instructions using

Note that when the -Kshortloop=N option is specified, performance may be reduced because of assumption that the iterationat the all innermost loops in the program is few.

noshortloop=N

-Knoshortloop directs that some optimizations are applied without assuming that the iteration at the all innermost loops is few.

- 42 -

{ simd[=level] | nosimd } level: { 1 | 2 | auto }

These options direct whether to create objects that use SIMD extension instructions. The -Ksimd option is effective only when theoptimization level -O2 option or higher is set. Specify 1, 2, or auto for level. The default value for level is auto. The default whenthe -O2 option or higher is set is the -Ksimd option.

simd=1

Objects that use SIMD extension instructions are created.

simd=2

In addition to the functions of the simd=1 option, objects that use SIMD extension instructions are created for loops that includeIF constructs. In order to redundantly execute instructions within IF constructs, performance may be reduced depending on thetrue ratio of the IF construct. In addition, as with the -Kpreex option, speculative executions are performed within an expressionin IF constructs, so this optimization may execute an instruction which would not be executed based on the logic of the program,causing an error. (*1)

simd=auto

The compiler automatically determines whether to use SIMD extension for the loop. SIMD extension is promoted for loopsthat contain IF construct. (*1)

When the -KHPC_ACE option is set, -Ksimd=1 is assumed.

nosimd

Objects are created that do not use SIMD extension instructions.

*1) When the -Knf option is effective, it is possible to avoid the error by speculative execution of load instructions in IF construct.See the explanation of the -Knf option in this section for notes when specifying the -Knf option.

{ simd_separate_stride | simd_noseparate_stride }

When the stride width of memory access in the loop of SIMD-target is out of the range of 1 to 7, the compiler determines whetherto use SIMD indirect instructions or scalar instructions at the memory access.

When the -O2 option or higher is effective and the -Ksimd option or the SIMD optimization control specifier is effective, the -Ksimd_separate_stride option is effective. The default is -Ksimd_noseparate_stride.

-Ksimd_separate_stride

When the stride width of memory access is out of the range of 1 to 7, the -Ksimd_separate_stride option specifies to use scalarinstructions at the memory access.

-Ksimd_noseparate_stride

The -Ksimd_noseparate_stride option specifies to use SIMD indirect instructions.

{ static_fjlib | nostatic_fjlib }

These options specify whether or not to create executable program which the Fortran intrinsic function library is linked in statically.These options must be set at linking. The -Knostatic_fjlib option is default.

static_fjlib

-Kstatic_fjlib option specifies to create the executable program which the Fortran intrinsic function library is linked in statically.The size of executable program may increase when specifying this option.

nostatic_fjlib

-Knostatic_fjlib option specifies to create the executable program which the Fortran intrinsic function library is linked indynamically.

{ striping[=N] | nostriping } 2 <= N <= 100

These options specify whether or not to perform the striping optimization. The striped length (number of expansions) can bespecified for N as a value between 2 and 100. If N is omitted, the compiler will automatically determine the best value. When theiteration count of a loop in the source program is apparent, the expansion number determined by the compiler will be used even ifa number that exceeds the iteration count is specified for N. The default is -Knostriping.

See Section "9.1.2.4 Loop Striping" for information on striping.

- 43 -

In striping, the statements in a loop are unrolled many times, so, as with loop unrolling, the size of the object module will increase,along with the required compilation time. Caution should be taken when using it, as performance may also be reduced due to anincreased number of registers used.

striping[=N]

The striping optimization is performed. -Kstriping is effective only when the -O2 option or higher is set.

nostriping

The striping optimization is not performed.

{ swp | noswp }

These options direct whether to perform the software pipelining optimization. However, if the effect of software pipelining is notexpected, software pipelining is not performed. When the -O0 or -O1 option is set, the default is -Knoswp. When the -O2 optionor higher is set, the default is -Kswp. See Section "9.1.1.6 Software Pipelining" for information on software pipelining.

swp

Software pipelining is performed. -Kswp is effective only when the -O2 option or higher is set.

noswp

Software pipelining is not performed.

swp_strong

The -Kswp_strong option specifies to perform software pipelining for more loops by easing its applicable condition.

This option may increase compilation time and memory requirement significantly.

Aside from this point, the functions and notes are the same as for the -Kswp option.

{ temparraystack | notemparraystack }

These options direct whether to allocate the following results of the stack area. The default is -Knotemparraystack.

- Intermediate results of the array operation

- Mask expression evaluation result when iteration count of DO CONCURRENT is constant.

temparraystack

Intermediate results of the array operation and mask expression evaluation result are allocated to the stack area. See Section"9.11 Effects of Allocating on Stack" for points to note when specifying this option.

notemparraystack

Intermediate results of the array operation and mask expression evaluation result are not allocated to the stack area.

{ threadsafe | nothreadsafe }

These options direct whether or not to create thread-safe object programs. The default is -Knothreadsafe.

threadsafe

Thread-safe object programs are created.

nothreadsafe

Created object programs not thread-safe.

{ unroll[=N] | nounroll } 2 <= N <= 100

These options direct that the loop unrolling optimization is performed. Specify the upper bound for loop unrolling in N, as a valuebetween 2 and 100. If a value was not specified for N, the compiler will automatically determine the best value. If the -O0 or -O1option is set, the default is -Knounroll. If the -O2 option or higher is set, the default is -Kunroll.

unroll[=N]

Loop unrolling is performed. -Kunroll is effective only when the -O1 option or higher is set.

nounroll

Loop unrolling is not performed.

- 44 -

{ uxsimd | nouxsimd }

These options specify whether or not to use SIMD extensions by UXSIMD optimization. The -Ksimd option and the -KHPC_ACEoption must be specified together. The -Knouxsimd option is default.

SIMD extensions are not used for intrinsic procedures. However, when the inline expansion is applied by optimization, the expandedinstructions become the target of SIMD extensions.

See Section "9.1.2.6 UXSIMD" for information about UXSIMD.

uxsimd

The -Kuxsimd option specifies to use SIMD extension by UXSIMD optimization.

nouxsimd

The -Knouxsimd option invalidates -Kuxsimd option.

visimpact

The -Kvisimpact option directs that the most appropriate automatically parallelized object for VISIMPACT (Virtual SingleProcessor by Integrated Multicore Parallel Architecture) is created. This performs the same optimization as when the -Kfast,paralleloption is specified. Using this optimization the numerical operations do not conform to IEEE 754.

vppocl

The -Kvppocl option directs that the optimization indicator NOVREC is treated equally to the optimization indicatorNORECURRENCE.

This option is effective only when the -Kocl option is set.

{ XFILL[=N] | NOXFILL } 1 <= N <= 100

These options direct that for array data to be written only in a loop, an instruction is created that allocates a cache line for writingto the cache without loading data from the memory.

XFILL instructions targeting the cache lines specified in N are generated.

Specify a value between 1 and 100 for N. If a value is not specified for N, the compiler will automatically determine a value. Thisoption is effective only when the -O2 option or higher is set. The default is -KNOXFILL. See Section "9.1.2.5 XFILL" forinformation on XFILL.

XFILL[=N]

XFILL instructions are output.

NOXFILL

XFILL instructions are not output.

-L directory

The -L option add directory to the list of directories in which the linker searches for libraries. This option and its argument are passedto the linker.

-M directory

The -M option specifies an alternate directory for module information files (.mod files).

.mod files are created in directory during compilation. For information about the compilation of modules, see "Chapter 10 FortranModules".

-N src_arg

src_arg: { { allextput | noallextput } | { alloc_assign | noalloc_assign } | { cancel_overtime_compilation |nocancel_overtime_compilation } | check_cache_arraysize | check_global | check_intrfunc | { coarray | nocoarray } |{ compdisp | nocompdisp } | { copyarg | nocopyarg } | { f90move | nof90move } | { freealloc | nofreealloc } | { hook_func |nohook_func } | { hook_time | nohook_time } | { line | noline } | lst[=lst_arg] | lst_out=file | { mallocfree | nomallocfree } |maxserious=maxnum | { obsfun | noobsfun } | { privatealloc | noprivatealloc } | quickdbg[=dbg_arg] | { recursive | norecursive }| { rt_tune | rt_notune } | rt_tune_func | rt_tune_loop[=kind] | { save | nosave } | { setvalue[=set_arg] | nosetvalue } | { use_rodata| nouse_rodata } | { Rtrap | Rnotrap } }

- 45 -

{ allextput | noallextput }

The -Nallextput option creates an external name which appears only on EXTERNAL statement. The -Nallextput option is default.

The -Nnoallextput option does not create an external name which appears only on EXTERNAL statement.

{ alloc_assign | noalloc_assign }

The -Nalloc_assign is specified, the allocatable assignment (Refer to Section "6.4.2 Allocatable Assignment" ) is processed inFortran 2003 standard when the -X9 or -X03 option is in effect. If the -X6 or -X7 is in effect, the -Nalloc_assign is not in effect. -Nnoalloc_assign is default.

alloc_assign

The -Nalloc_assign is specified, the allocatable assignment is processed in Fortran 2003 standard when the -X9 or -X03 optionis in effect. The program has less effect on execution performance in Fortran 95 standard program when the -Nalloc_assign isspecified.

noalloc_assign

The -Nnoalloc_assign is specified, the allocatable assignment is not processed in Fortran 2003 standard.

{ cancel_overtime_compilation | nocancel_overtime_compilation }

These options specify whether to cancel the compilation if the compiler forecasts that it takes a long time (24 hours or more as aguide) to compile the program.

The -Ncancel_overtime_compilation option is default.

cancel_overtime_compilation

The -Ncancel_overtime_compilation specifies to cancel the compilation if the compiler forecasts that it takes a long time tocompile the program.

nocancel_overtime_compilation

The -Nnocancel_overtime_compilation specifies not to cancel the compilation even if it takes a long time to compile theprogram.

check_cache_arraysize

The -Ncheck_cache_arraysize options specifies to check the sizes of arrays during compilation and issue level-i diagnostic messagesif the array sizes may cause execution performance decrease.

This function decides that an array could be a cause of a cache conflict if its size is a multiple of that of second level cache.

The array shape could be rewritten so as not to have a size of multiple of that of the cache in order to avoid the impact.

The function cannot judge correctly when the sector cache is active.

When the compiler option -Karraypad_const[=N] or -Karraypad_expr=N is specified for notes when this function is used, the sizeof the array becomes the size which adds the size of the padding.

See Section "4.1.1 Compilation Diagnostic Messages" for information about diagnostic messages output during compilation.

check_global

The -Ncheck_global option specifies to check the size of common blocks, and the procedure characteristics between externalprocedure definitions and references, and between external procedure definitions and interface body in program units duringcompilation.

If an error occurs, a diagnostic message is output. When -Ncheck_global is valid, compilation time may increase.

check_intrfunc

The -Ncheck_intrfunc option specifies to make the intrinsic function error processing effective for object programs.

If this option is in effect, error processing is performed when the object program is executed.

When it suits the error as the result of checking, when the following processing is executed:

- The output of the diagnostic message, traceback and error summaries

- user control of error by ERRSET and ERRSAV subroutine

- 46 -

- execution of user-defined subroutine

See Section "8.1.5 Intrinsic Function Error Processing" for details about the error number is detected by -Ncheck_intrfunc optionand intrinsic function error processing.

{ coarray | nocoarray }

These options direct whether to enable COARRAY specifications. See the "Fortran User's Guide Additional Volume COARRAY"for information on this function. The default is -Nnocoarray.

coarray

The -Ncoarray option validates COARRAY specification.

Even when only object programs are specified as file names (that is, when only linking is performed), -Ncoarray must bespecified if an object program that was compiled by with the -Ncoarray option specified is included.

nocoarray

The -Nnocoarray option invalidates COARRAY specification.

If you compile COARRAY programs with this option, error messages are output and the compilation stops.

{ compdisp | nocompdisp }

These options specify whether or not to output file name and program name which are currently compiled. The -Nnocompdispoption is default.

{ copyarg | nocopyarg }

The -Ncopyarg option specify to copy scalar constant argument to generated variable argument. When the -Ncopyarg is specified,even if the value of dummy argument is updated in procedure, it doesn't affected by the constant value of actual argument.

The program is not Fortran standard compliant. However, the program that can be operated by specifying this option does notconform to with the Fortran 95 standard.

Example:

CALL SUB(1)PRINT *,1ENDSUBROUTINE SUB(I)I=2END

When -Ncopyarg is specified, this program outputs 1.

{ f90move | nof90move }

The -Nf90move option is specified, the character assignment statements are processed in Fortran standard even if the -X7 or -X6option is in effect.

The -Nnof90move option is specified, the character assignment statements are not processed in Fortran standard.

If the -X9 or -X03 option is in effect, the -Nf90move is always in effect and the -Nnof90move option cannot be specified. If the -X7 or -X6 option is in effect, the -Nnof90move option is default.

Example:

CHARACTER(LEN=5) CC='12345'C(2:5)=C(1:4)PRINT *,C ! output is 11234END

When the -Nf90move option is specified, this program outputs "11234" even if the -X7 or -X6 option is in effect.

{ freealloc | nofreealloc }

The -Nfreealloc option specifies to deallocate arrays which do not have save attribute and are allocated when the procedure exits.The -Nfreealloc option is default.

- 47 -

If -Nnofreealloc option is specified and an unsaved allocatable array has a status of currently allocated when a procedure is exited,the array is not deallocated. This is not the behavior specified by the Fortran 95 standard.

{ hook_func | nohook_func }

These options specify whether or not to use the hook function that is called from a specified location.

See Section "8.2.3 Hook Function", for information about the hook function.

The default is -Nnohook_func.

This option must be set if an object program compiled with the -Nhook_func option set is included, even if only the object programname appears (that is, if only linking is performed).

hook_func

The -Nhook_func option is specified, a user-defined subroutine is called from the following locations:

- Program entry and exit

- Procedure entry and exit

- Parallel region (OpenMP or automatic parallelization) entry and exit

nohook_func

This option specifies not to use the hook function.

{ hook_time | nohook_time }

These options specify whether or not to use the hook function called at regular time interval.

Refer to Section "8.2.3 Hook Function", for information about the hook function.

The default is -Nnohook_time. This option must be set at linking.

hook_time

If the -Nhook_time option is enabled, a user-defined subroutine is called at regular time interval.

The time interval can be specified using the environment variable FLIB_HOOK_TIME. If the environment variableFLIB_HOOK_TIME is not specified, the user-defined subroutine is called once every minute.

Refer to Section "3.8 Using Environment Variables for Execution", for information about the environment variableFLIB_HOOK_TIME.

nohook_time

This option specifies not to use the hook function.

{ line | noline }

The -Nline option specifies to output the statement number where the user defined procedure which caused an error is called andthe statement number where an error occurred in the procedure. These numbers are output into the trace back map. The -Nlineoption is default.

The -Nnoline option specifies that even if the error occurs in the procedure, the statement number is not displayed in the trace backmap.

This information is output to standard error output when an error occurs during execution of Fortran programs. See Section "4.2.2Trace Back Map for information about the trace back map.

lst[=lst_arg] lst_arg:{a|d|i|m|p|t|x}

The function of this option is equivalent to the -Q{a|d|i|m|p|t|x} option. For more details about the arguments, see -Q option.

Note that usage of comma to specify multiple arguments is different from the -Q option.

Each argument delimited by comma should be started from "lst=".

Example: "-Nlst=a,lst=d,lst=x"

lst_out=file

The function of this option is equivalent to the -Qofile option.

The compilation information is output to the file.

- 48 -

{ mallocfree | nomallocfree }

These options specify whether or not to evaluate MALLOC and FREE as intrinsic procedures. The -Nnomallocfree is default.

-Nmallocfree is specified to evaluate MALLOC and FREE as intrinsic procedures.

When -Nnomallocfree option is specified, evaluates MALLOC and FREE as service routines.

maxserious=maxnum

Stop the compilation if s-level (serious) errors are detected more than maxnum times. maxnum must be greater than or equal to 1.When this option is omitted, compilation does not stop even if s-level (serious) errors are detected.

{ obsfun | noobsfun }

The -Nobsfun option specifies that the under mentioned names are interpreted as intrinsic functions.

The -Nnoobsfun option specifies that the under mentioned names are not interpreted as intrinsic functions. The -Nnoobsfun optionis default.

AIMAX0, AJMAX0, I2MAX0, IMAX0, JMAX0, IMAX1, JMAX1, AIMIN0, AJMIN0,I2MIN0, IMIN0, JMIN0, IMIN1, JMIN1, FLOATI, FLOATJ, DFLOTI,DFLOTJ,IIABS, JIABS, I2ABS, IIDIM, JIDIM, I2DIM,IIFIX, JIFIX, JFIX,INT1, INT2, INT4, INT8, IINT, JINT,ININT, JNINT, IIDNNT, I2NINT,JIDNNT, IIDINT, JIDINT,IMOD, JMOD, I2MOD, IISIGN, JISIGN,I2SIGN,BITEST, BJTEST, IIBCLR, JIBCLR, IIBITS, JIBITS,

IIBSET, JIBSET,IBCHNG, ISHA, ISHC, ISHL,IIAND, JIAND, IIEOR, JIEOR,IIOR, JIOR, INOT, JNOT,IISHFT, JISHFT, IISHFTC, JISHFTC,IZEXT, JZEXT,IZEXT2, JZEXT2, JZEXT4,VAL

{ privatealloc | noprivatealloc }

For a list item with the ALLOCATABLE attribute in a private clause of OpenMP: if the list item is "currently allocated", the newlist item will have an initial state of "not currently allocated". If the -Nprivatealloc option is specified, the program does not conformto the OpenMP 3.1 standard. The -Kopenmp option must be specified together.

When the -Nnoprivatealloc option is specified, the -Nprivatealloc option is ignored. The -Nnoprivatealloc option is default.

quickdbg[=dbg_arg] dbg_arg: {{ argchk | noargchk } | { subchk | nosubchk } | { undef | undefnan | noundef } | { inf_detail| inf_simple }}

This function is used to debug Fortran source code. If the debug function is enabled, it checks for Fortran source errors duringcompilation. It also embeds in object programs the information required for debugging and performs automatic checks duringexecution.

Checks are performed if the -Nquickdbg=argchk, -Nquickdbg=subchk, -Nquickdbg=undef, or -Nquickdbg=undefnan options areset.

For dbg_arg, either argchk, noargchk, subchk, nosubchk, undef, undefnan, noundef, inf_detail, or inf_simple can be specified.These options can be used in combination by specifying the -Nquickdbg[=dbg_arg] option multiple times.

The -Nquickdbg=inf_detail and -Nquickdbg=inf_simple options are enabled when set together with either the -Nquickdbg, -Nquickdbg=argchk, -Nquickdbg=subchk, -Nquickdbg=undef, or -Nquickdbg=undefnan option.

Omitting dbg_arg is equivalent to specifying -Nquickdbg=argchk,fastdbg=subchk,fastdbg=undef.

If the -Nquickdbg=argchk, -Nquickdbg=subchk, -Nquickdbg=undef, or the -Nquickdbg=undefnan option is enabled, the -Nquickdbg=inf_detail and -Ncheck_global options are also enabled.

See "8.2 Debugging Functions" for detail.

If the -H option is enabled, disable the -Nquickdbg option.

{ argchk | noargchk }

This option specifies whether or not to check the validity of the called and calling of procedure references.

This check is performed at compilation and at execution. See Section "8.2.1.1 Checking Argument Validity (ARGCHK)" fordetail.

- 49 -

-Nquickdbg=argchk

Checks the validity of arguments and result of procedure. If an error is detected, a diagnostic message is output.

-Nquickdbg=noargchk

The validity of arguments and result of procedure are not checked.

{ subchk | nosubchk }

This option specifies whether or not the validity of the declared expression and referenced expression are checked in arraysection, array element, and substring declarations and references.

This check is performed at execution. See Section "8.2.1.2 Checking Subscript and Substring Values (SUBCHK)" for details.

-Nquickdbg=subchk

Checks the subscript and substring value. If an error is detected, a diagnostic message is output.

-Nquickdbg=nosubchk

The subscript and substring value are not checked.

{ undef | undefnan | noundef }

This option specifies whether or not to check that data is defined in references to module declaration parts and variables outsideof common blocks.

This check is performed when the program is executed. See Section "8.2.1.3 Checking References for Undefined Data Items(UNDEF)" for details.

-Nquickdbg=undef

Checks undefined data references. If an error is detected, a diagnostic message is output.

-Nquickdbg=undefnan

Checks undefined data references. When this option is specified, real type and complex type errors are detected as invalidoperation exceptions. When an error is detected, either a diagnostic message or an invalid operation exception message is output,depending on the variable type.

If both the -Nquickdbg=undefnan option and the -Kfed option are specified, the one specified last is valid.

-Nquickdbg=noundef

Undefined data reference checks are not performed.

{ inf_detail | inf_simple }

This option specifies the information to be included in diagnostic messages output when errors are detected.

See Section "8.2.1 Debugging Check Functions" for details.

-Nquickdbg=inf_detail

In addition to the message and line number where the error occurred, information is output such as the variable name, procedurename, and element location in order to identify the cause.

However, when -Nquickdbg=undefnan option is specified, for errors detected as invalid operation exceptions, only the errorand the line number where it occurred are output in diagnostic messages.

-Nquickdbg=inf_simple

The error and the line number where it error occurred are output in diagnostic messages.

{ recursive | norecursive }

These options are specified, the RECURSIVE keyword is added in each SUBROUTINE and FUNCTION statement.

Example:

SUBROUTINE SUB

is evaluated as follows if the -Nrecursive option is specified.

- 50 -

RECURSIVE SUBROUTINE SUB

When -Nrecursive option is effective, the program does not conform to the Fortran 95 standard.

When -Nnorecursive option is specified, the RECURSIVE keyword is not added in SUBROUTINE and FUNCTION statements.The -Nnorecursive option is default.

{ rt_tune | rt_notune }

The -Nrt_tune option specifies to output the runtime information (parallelization information, cost information, input-outputinformation and hardware monitor information).

The runtime information can be used for the tuning of user program. See "Runtime Information Output Function User's Guide" fordetails.

When the -Nrt_notune option is specified, the -Nrt_tune option is ignored. The -Nrt_notune option is default.

rt_tune_func

In addition to the -Nrt_tune output, runtime information about a user-defined functions is output.

This option induces the -Nrt_tune option. When the -Nrt_notune option is specified after -Nrt_tune_func option, -Nrt_tune_funcoption will be invalid.

See "Runtime Information Output Function User's Guide" for details.

rt_tune_loop[=kind] kind: { all | innermost }

In addition to the -Nrt_tune output, runtime information about loops is output.

All or innermost can be specified in the argument kind. If the kind is not specified, all is used.

This option induces the -Nrt_tune option. When the -Nrt_notune option is specified after -Nrt_tune_loop option, -Nrt_tune_loopoption will be invalid.

See "Runtime Information Output Function User's Guide" for details.

rt_tune_loop=all

Runtime information about all loops is output.

rt_tune_loop=innermost

Runtime information about the innermost loops is output.

{ save | nosave }

When -Nsave option is specified, the SAVE statement without saved entity list is added in each program unit except main program.It is ignored if the -Nrecursive or -Kauto option is specified.

When -Nnosave option is specified, The SAVE statement without saved entity list is not added in program unit. The -Nnosaveoption is default.

{ setvalue[=set_arg] | nosetvalue }set_arg: { { heap | noheap } | { stack | nostack } | { scalar | noscalar } | { array | noarray } | { struct | nostruct } }

The options specify whether or not to automatically set zero value by the procedure entrance and ALLOCATE statement in variablesthat are allocated to stack or heap. The -Nnosetvalue is default.

Note that execution time may increase for the initialization.

If the -Eg, -H or -Nquickdbg option is specified, the -Nsetvalue option is ignored.

setvalue[=set_arg]

The -Nsetvalue option specifies to automatically set zero value in variables.

Either heap, noheap, stack, nostack, scalar, noscalar, array, noarray, struct or nostruct can be specified in the set_arg.

These options can be used in combination by specifying the -Nsetvalue=set_arg option multiple times.

If the set_arg is omitted, it is equivalent to the following options:

-Nsetvalue=heap,setvalue=stack,setvalue=scalar,setvalue=array,setvalue=struct

- 51 -

{ heap | noheap }

These options specify whether or not to automatically set zero value by the procedure entrance and ALLOCATE statement invariables that are allocated to heap.

The zero value is not set to the allocated area by service subroutine MALLOC and intrinsic function MALLOC. The zero valueis not set in the allocatable variable at head of the OpenMP directive block when the allocatable variable is specified byPRIVATE or LASTPRIVATE clause of OpenMP directive.

The value is set in the following variables:

- Automatic data objects when -Knoautoobjstack option is effective.

- Allocatable variables.

- Pointer variables.

setvalue=heap

The -Nsetvalue=heap option specifies to automatically set the zero value by the procedure entrance and ALLOCATE statementin variables that are allocated to heap.

When the setvalue=heap is specified, the -Nsetvalue=scalar, -Nsetvalue=array and -Nsetvalue=struct options are in effect.

setvalue=noheap

The -Nsetvalue=noheap option does not apply to automatically set the zero value by the procedure entrance and ALLOCATEstatement in variables that are allocated to heap.

{ stack | nostack }

These options specify whether or not to automatically set the zero value by the procedure entrance in variables that allocatedto stack.

The value is set in the following variables:

- Automatic data objects when -Kautoobjstack option is effective.

- Local variables without SAVE attribute when -Kauto option is effective.

- Local variables in program unit that has RECURSIVE or PURE.

- Local variables that declare by AUTOMATIC attribute or AUTOMATIC statement.

- Local variables that specify by PRIVATE clause or LASTPRIVATE clause in OpenMP directive.

setvalue=stack

The -Nsetvalue=stack option specifies to automatically set the zero value by the procedure entrance in variables that are allocatedto stack.

But, the zero value is set at head of the OpenMP directive block when the specifying variables by PRIVATE clause orLASTPRIVATE clause.

When the setvalue=stack is specified, the -Nsetvalue=scalar, -Nsetvalue=array and -Nsetvalue=struct options are in effect.

setvalue=nostack

The -Nsetvalue=nostack option does not apply to automatically set the zero value by the procedure entrance in variables thatare allocated to stack.

{ scalar | noscalar }

These options specify whether or not to automatically set the zero value in scalar variables of numeric types and logical types.

setvalue=scalar

The -Nsetvalue=scalar option specifies to automatically set the zero value in scalar variables of numeric types and logical types.

setvalue=noscalar

The -Nsetvalue=noscalar option does not apply to automatically set the zero value in scalar variables of numeric types andlogical types.

- 52 -

{ array | noarray }

These options specify whether or not to automatically set the zero value in array variables of numeric types and logical types.

setvalue=array

The -Nsetvalue=array option specifies to automatically set the zero value in array variables of numeric types and logical types.

setvalue=noarray

The -Nsetvalue=noarray option does not apply to automatically set the zero value in array variables of numeric types and logicaltypes.

{ struct | nostruct }

These options specify whether or not to automatically set the zero value in scalar and array variables of derived type andcharacter type.

setvalue=struct

The -Nsetvalue=struct option specifies to automatically set the zero value in scalar and array variables of derived type andcharacter type.

The variable of derived type having length type parameter is not applied.

setvalue=nostruct

The -Nsetvalue=nostruct option does not apply to automatically set the zero value in scalar and array variables of derived typeand character type.

nosetvalue

The -Nnosetvalue option does not apply to automatically set zero value in variables.

The option is equivalent to the following options:

-Nsetvalue=noheap,setvalue=nostack,setvalue=noscalar,setvalue=noarray,setvalue=nostruct

Be careful of the following:

1. When the -Nsetvalue option is in effect, it is necessary to be the following conditions:

- The -Nsetvalue=heap or -Nsetvalue=stack is in effect, and

- The -Nsetvalue=scalar, -Nsetvalue=array or -Nsetvalue=struct is in effect.

2. If the -Nsetvalue=heap or -Nsetvalue=stack is specified, -Nsetvalue=scalar, -Nsetvalue=array and -Nsetvalue=struct aregenerated. Therefore, when the -Nsetvalue=noscalar, -Nsetvalue=noarray or -Nsetvalue=nostruct will be in effect, the optionshould be specified after -Nsetvalue=heap or -Nsetvalue=stack.

3. When the zero value is set in only scalar variable of numeric types and logical types for stack, it is necessary to be the followingoptions:-Nsetvalue=stack,setvalue=noarray,setvalue=nostruct

When the -Nsetvalue=stack is specified, the -Nsetvalue=scalar, -Nsetvalue=array and -Nsetvalue=struct are generated.Therefore, if -Nsetvalue=stack,setvalue=scalar are specified, the option specifying is not limited to scalar variables alone.

{ use_rodata | nouse_rodata }

The -Nuse_rodata option is specified, string constant, floating point constant and initialization value of aggregate type local storagevariable are allocated to read-only data section. The -Nuse_rodata option is default.

When the -Nnouse_rodata option is specified, the -Nuse_rodata option is ignored.

{ Rtrap | Rnotrap }

The -NRtrap option specifies to trap the intrinsic instructions errors and floating point exceptions in execution, and if the intrinsicinstructions errors or floating point exceptions are trapped, the runtime messages are outputted. -NRnotrap option is the default.This option is not effective in an initial constant expression that is operated at compilation time.

Rtrap

If a main program unit is compiled with -NRtrap option, the intrinsic instructions errors (from jwe0259i-e to jwe0282i-e,jwe1397i-e, jwe1398i-e, jwe1413i-e, jwe1414i-e, jwe1416i-e, and jwe1417i-e) and the floating point exceptions (jwe0011i-u,

- 53 -

jwe0012i-u, jwe0013i-u and jwe0292i-u) are trapped and runtime messages are output. The floating-point underflow exceptionis trapped when the environment variable FLIB_EXCEPT=u have been specified.

If both the -NRtrap option and the -Kfed option are specified, the one specified last is valid.

If -NRtrap option is used together with -Kpreex option and -Ksimd=2 option, the speculative execution may cause exceptionsthat would not normally occur.

For the messages, see "Fortran/C/C++ Runtime Messages" for details.

Rnotrap

If a main program is compiled with -NRnotrap option, the intrinsic instructions errors and the floating point exceptions are nottrapped and runtime messages are not output.

-O [ opt_lvl ] opt_lvl: { 0 | 1 | 2 | 3 }

The -O option specifies the optimization level used by the compiler.

The system provides four optimization levels: 0, 1, 2 and 3. If the argument is omitted, level 3 is used. If the -O option is not specified,level 2 is used.

See "Chapter 9 Optimization Functions" for information about the optimization of object programs.

In addition, the following optimization facilities can be specified:

- The advance evaluation of invariant expressions and the evaluation strategy can be changed by specifying -K series option

- It can indicate the inline expansion of user-defined procedure by specifying -x option.

0

The -O0 option creates an object program without applying optimizations. A program compiled with the -O0 option requires theleast compilation time and memory. Specify this argument to debug compilation errors in Fortran source programs.

1

The -O1 option creates an object program by applying basic optimization. The run time of an executable program will be shorterand the size of it will be smaller if it is created with the -O1 option than if it is created with the -O0 option.

2

When the -O2 option is effective, following optimizations are applied in addition to the optimizations applied by the -O1 option.

- Loop unrolling

- Software pipelining

- Loop blocking

- Loop fusion

- Loop fission

- Exchange Loop (-Kloop_interchange)

- Using prefetch instructions( Equivalents to -Kprefetch_sequencial, prefetch_cachelevel=all )

- Using SIMD instructions

- Repeat basic optimization

The -O2 option repeats optimizations applied by the -O1 option until it admits of no optimization.

Note that compile time and object program size may be increase compared to the -O1 option. To determine whether eachoptimization was applied, specify the -Koptmsg=2 option.

3

When the -O3 option is effective, following optimizations are applied in addition to the optimizations applied by the -O2 option.

- Unrolling nested loops

- Splitting for promoting loop exchange

- 54 -

- Loop unswitching

Note that compile time and object program size may be increase compared to the -O2 option.

To determine whether each optimization was applied, specify the -Koptmsg=2 option.

-P

The -P option specifies to perform only preprocessing to the specified source file. The result of preprocessing is output to the temporaryfile.

The -Cpp option must be specified together unless a suffix of the source file is .F, .FOR, .F90, .F95 or .F03. The -E option cannot bespecified together.

If this option is set, only preprocessor is performed. This option cannot be specified with -E.

The temporary files are generated are as follows:

Source file Result file of preprocessing

file.Ffile.cpp.f

file.f

file.FORfile.cpp.for

file.for

file.F90file.cpp.f90

file.f90


file.f95


file.f03

-Q [ lst_arg ] lst_arg: { a | d | i | m | ofile | p | t | x }

The -Q option outputs compilation information to a file with suffix .lst. If more than one Fortran source program files is specified, itis output to the first-file-name.lst.

The -Q option can be specified with arguments a, d, i, m, ofile, p, t, and x. The arguments may be specified at the same time, separatedby comma as in -Qa,ofile,x. If arguments are omitted, the source program list and diagnostic error messages are output.

a

If the -Qa option is specified, the attributes of names are output in addition to the compilation information produced by the -Qoption.

The function of this option is equivalent to the -Nlst=a option.

d

If the -Qd option is specified, the layout of derived types is output in addition to the compilation information produced by the -Qoption.

The function of this option is equivalent to the -Nlst=d option.

i

If the -Qi option is specified, include files are listed in addition to the compilation information produced by the -Q option.

The function of this option is equivalent to the -Nlst=i option.

m

If the -Qm option is specified, the Fortran program which expresses the situation of automatic parallelization with OpenMPdirectives is output in addition to the compilation information produced by the -Qp option. Fortran program to be output is storedin the file having name that .omp is inserted before suffix (.f, .F, etc.) of an input file. It is ignored if the -Qi option is specified, orif -Kparallel option is not effective.

- 55 -

The function of this option is equivalent to the -Nlst=m option.

ofile

The file name must be specified immediately after o. The compilation information is output to the file.

The function of this option is equivalent to the -Nlst_out=file option.

p

The -Qp option specifies optimization information to be output along with listing of -Q option. Optimization information showsthe situation of automatic parallelization, inlining and unrolling.

The function of this option is equivalent to the -Nlst=p option.

t

If the -Qt option is specified, the details optimization information and the statistics information is output in addition to thecompilation produced by the -Qp option.

The function of this option is equivalent to the -Nlst=t option.

x

If the -Qx option is specified, the cross reference list of name and label is output in addition to the compilation produced by the -Q option.

The function of this option is equivalent to the -Nlst=x option.

-S

The -S option suppresses creation of the corresponding object programs and does not call the linker. Instead, assembly programs (fileswith the suffix .s) are created. If only assembly source files or object files are specified by file, it is meaningless to invoke the compilecommand. If the -S option is effective, the -Klto option will be invalid.

{ -SSL2 | -SSL2BLAMP }

These options relate to linking with Fujitsu's math libraries (SSL II, BLAS, LAPACK). The details are described in "SSL II OnlineDocuments", "SSL II Thread-Parallel Capabilities Online Documents" and "BLAS, LAPACK, ScaLAPACK Online Documents."Following is just an outline.

-SSL2

The whole set of routines from SSL II, SSL II Thread-Parallel Capabilities and BLAS/LAPACK becomes part of link libraries.

-SSL2BLAMP

The whole set of routines from SSL II, SSL II Thread-Parallel Capabilities and BLAS/LAPACK Thread-Parallel versions becomespart of link libraries (i.e. -SSL2BLAMP just replaces the sequential BLAS/LAPACK from -SSL2 with the corresponding thread-parallel versions.)

-U name

The -U option undefines name, which has the same effect as an #undef preprocessing directive. If the same name is specified for both-D and -U, name is not defined regardless of the order of the options.

-V

The -V option displays the version and release information on each Fortran compiler component.

-W tool,arg1[,arg2]...

Passes each argument arg1[,arg2]... as a separate argument to a tool. The arguments must be separated by commas (A comma can bepart of an argument by using a backslash as an escape character before it. The backslash is removed from the resulting argument).

tool can be one of the following:

Value Meaning

p Preprocessor

0 Compiler

a Assembler

- 56 -

Value Meaning

l Linker

For example, -Wa,-oobjfile passes -o and objfile to the assembler, in that order. Also, -Wl,-Iname causes the linking phase to overridethe default name of the dynamic linker. For other -W options, arguments passed to a tool may not be in order.

If the -Wa option is specified, the -Klto option will be invalid.

-X lan_lvl lan_lvl: { 6 | 7 | 9 | 03 | d7 }

The -X option directs the level of language specification. Either 6, 7, 9, 03 or d7 can be specified as the argument of the -X option.

The interpretation of Fortran source programs is different for each level of language specification. To compile Fortran source programs,indicate the level of language specification used by the Fortran compiler. See Section "6.1 Fortran Versions" for information on thedifference between languages.

6

The -X6 option specifies to compile Fortran source programs as FORTRAN66 source.

7

The -X7 option specifies to compile Fortran source programs as FORTRAN77 source. If the suffix of the Fortran source file is .for .F, -X7 is the default.

9

The -X9 option specifies to compile Fortran source programs as Fortran 95 or Fortran 90 source. Activates -Nf90move option atthe same time. If the suffix of Fortran source file is .f90, .F90, .f95 or .F95, -X9 is the default.

03

The -X03 option specifies to compile Fortran source programs as Fortran 2003 source.

Activates -Nf90move option at the same time. If the suffix of Fortran source file is .f03 or .F03, -X03 is the default.

d7

If the version of the language is not specified, and the suffix of the Fortran source file is .f, .for, .F or .FOR, -Xd7 is the default.

-#

The -# option specifies to output the name of each path and options used by frtpx command, but does not execute any of the tools.

-###

The -### option specifies to output the name of each path and options used by frtpx command as it executes.

2.3 Compile Command Environment VariableIt explains the environment variables that the compile commands (frtpx command and frt command) recognize as follows.

FORT90C

This is the environment variable that the frt command (own compiler) recognizes.

The compiler options specified in the environment variable FORT90C are combined with the frt command option arguments. The frtcommand option arguments have precedence if there is a conflict with options specified in the environment variable FORT90C.

An example to use the environment variable FORT90C is shown in the following.

$ export FORT90C="-fw -I/usr/prv/usri"$ frt a.f90

This is interpreted as

$ frt -fw -I/usr/prv/usri a.f90

It explains the environment variable that the frtpx command (cross compile command) recognizes as follows.

- 57 -

FORT90CPX

This is the environment variable that the frtpx command (corss compiler) recognizes.

The compiler options specified in the environment variable FORT90CPX are combined with the frtpx command option arguments.The frtpx command option arguments have precedence if there is a conflict with options specified in the environment variableFORT90CPX.

An example to use the environment variable FORT90CPX is shown in the following.

$ export FORT90CPX="-fw -I/usr/prv/usri"$ frtpx a.f90

This is interpreted as

$ frtpx -fw -I/usr/prv/usri a.f90

TMPDIR

This is the environment variable that the frtpx command (corss compiler) and the frt command (own compiler) recognize.

The temporary directory used by the frtpx command or frt command can be changed by specifying a value for the environment variableTMPDIR.

If the environment variable TMPDIR is not set, /tmp is used. To avoid to output in common directory, set the environment variableTMPDIR to local directory.

An example to use the environment variable TMPDIR is shown in the following.

$ export TMPDIR=/usr/local/tmp

2.4 Compilation Profile FileThe default of compilation options can be changed by specifying the compilation profile file (/etc/opt/FJSVmxlang/jwd_prof).

A compilation profile file must be used according to the following format:

- Blank characters not enclosed quotation mark (") or single quotation mark (') are not significant.

- Either single or double quotation marks can be used to begin and end character strings. A character string is terminated be the samedelimiter with which the string began. A character string is also terminated by the end of a line.

- A symbol (#) that is not part of a character string starts a comment. Any characters from the # to the end of the line are treated as partof the comment.

The following is an example of a compilation profile file specification:

#Default options-fW -Kocl

The priority of compilation options is as follows, highest first:

1. Compiler directive lines(the -Koptions is specified)

2. Compilation command operand

3. Environment variable FORT90CPX or FORT90C

4. Profile file

5. Default value

2.5 Compiler Directive LinesIf !options is specified the first in Fortran source program and compiler option -Koptions is effective, options in !options become effective.

Form of compiler directive lines:

- 58 -

!options opt[,opt] ....

opt: compiler options

List of effective compiler options:

-O[1-3]

-Keval/noeval/loop_blocking[=N]/loop_noblocking/loop_fusion/loop_nofusion/ilfunc/noilfunc/mfunc[=level]/nomfunc/ocl/noocl/preex/nopreex/noprefetch/prefetch_infer/prefetch_noinfer/prefetch_iteration=N/prefetch_iteration_L2=N/prefetch_line=N/prefetch_line_L2=N/striping[=N]/nostriping/unroll[=N]/nounroll/noalias[=spec]/vppocl

Example:

1---5----10---15---20---25!options -O2 -Knounroll subroutine sub

Notes:

- !options cannot be specified the first in module procedure or internal subprogram.

- If compiler option -O3 effective, -O[1|2] option cannot be specified in compiler directive lines.

2.6 Return Values during CompilationThe following table lists the return values set by the frtpx(1) command.

Return value Status

0 Normal termination

1 Compile or linker error

Other than the above Linker error

- 59 -

Chapter 3 Executing Fortran ProgramsThis chapter describes the procedures for executing Fortran programs.

3.1 Execution CommandAn execution command is used to execute the executable program file created by the frtpx compile command.

3.2 Execution Command FormatRuntime options and user-defined executable program options may be specified as command option arguments of an execution command.The runtime options use functions supported by the Fortran library.

See Section "3.3 Runtime Options", for information on the runtime option.

The format of runtime options is as follows:

exe_file [-Wl,-runtime option[,-runtime option]...] [-user-defined executable program option[,-user-defined executable program option]...]

Notes:

1. exe_file indicates the executable program file.

2. The GETPARM and GETARG service subroutines fetch user-defined executable program options other than runtime options.

3. If an option is specified more than once with different arguments, the last occurrence is used.

The following example shows how to specify the runtime option -Wl,-i and user-defined executable program options -K and -L as commandoption arguments of a command named a.out.

$ ./a.out -Wl,-i -K,-L

3.3 Runtime OptionsRuntime options may be specified as arguments for execution commands or in the FORT90L environment variable. This section explainsthe format and functions of the runtime options.

The runtime option format is as follows:

-Wl[,-a] [,-dnum] [,-enum] [,-gnum] [,-i] [,-lelvl] [,-mu_no] [,-n] [,-pu_no] [,-q] [,-ru_no] [,-tsec] [,-u] [,-x] [,-Cu_no] [,-Gu_no] [,-Lb] [,-Li] [,-Lr] [,-Lu] [,-M] [,-Q] [,-Re] [,-Rp] [,-Ry] [,-Tu_no]

When runtime options are specified, -Wl (l is lowercase) is required at the beginning of the runtime options, and the options must beseparated by commas. If the same runtime option is specified more than once with different arguments, the last occurrence is used.

Example: Runtime option

$ ./a.out -Wl,-a,-p10,-x

-a

When the -a option is specified, an abend is executed forcibly following normal program termination. This processing is executedimmediately before closing external files.

Example: Runtime option -a

$ ./a.out -Wl,-a

- 60 -

-dnum 1 <= num <= 32767

The -d option determines the size of the input/output buffer used by a direct access input/output statement. The -d option improvesinput/output performance when data is read from or written to files a record at a time in sequential record-number order. If the -d optionis specified, the input/output buffer size is used for all units used during execution.

To specify the size of the input/output buffer for individual units, specify the number of Fortran records in the environment variablefuxxbf (xx is the unit number). See Section "3.8 Using Environment Variables for Execution" for details. When the -d option and theenvironment variable are specified at the same time, the environment variable becomes effective. The option argument num specifiesthe number of Fortran records, in fixed-block format, included in one block. The option argument num must be an integer from 1 to32767. The other integer value is specified, the -d option is invalid. To obtain the input/output buffer size, multiply num by the valuespecified in the RECL= specifier of the OPEN statement. If the files are shared by several processes, the number of Fortran recordsper block must be 1. If the -d option is omitted, the size of the input/output buffer is 8M bytes.

Example: Runtime option -d

$ ./a.out -Wl,-d10

-enum 0 <= num <= 32767

The -e option controls termination based on the total number of execution errors. The option argument num, specifies the error limitas an integer from 0 to 32767. When num is greater than or equal to 1, execution terminates when the total number of errors reachesthe limit. If -enum is omitted or num is zero, execution is not terminated based on the error limit. However, program execution stillterminates if the Fortran system error limit is reached.

Example: Runtime option -e

$ ./a.out -Wl,-e10

-gnum 0 <= num <= 2097151

The -g option sets the size of the input/output buffer used by a sequential access input/output statement. This size is set in units ofkilobytes for all unit numbers used during execution. The argument num must be an integer from 1 to 2097151. The other value isspecified, the -g option is invalid. If the -g option is omitted, the size of the input/output defaults to 8M bytes.

The -g option improves input/output performance when a large amount of data is read from or written to files by an unformattedsequential access input/output statement. The argument num is used as the size of the input/output buffer for all units. To avoid usingexcessive memory, specify the size of the input/output buffer for individual units by specifying the size in the environment variablefuxxbf (xx is the unit number). See Section "3.8 Using Environment Variables for Execution" for details. When the -g option is andthe environment variable are specified at the same time, the environment variable becomes effective.

Example: Runtime option -g

$ ./a.out -Wl,-g10

-i

The -i option controls processing of runtime interrupts. When the -i option is specified, the Fortran library is not used to processinterrupts. When the -i option is not specified, the Fortran library is used to process interrupts. Diagnostic messages are output. Forthe error check, see Section "8.2.2 Debugging Programs for Abend" for details.

For the interrupt, see Section "8.1.7 Exception Handling Processing" for details.

Example: Runtime option -i

$ ./a.out -Wl,-i

-lelvl elvl: { i | w | e | s }

The -l option controls the output of diagnostic messages during execution. The option argument elvl, specifies the lowest error level,i, w, e, or s, for which diagnostic messages are to be output. If the -l option is not specified, diagnostic messages are output for errorlevels w, e, and s. However, messages beyond the print limit are not printed.

Example: Runtime option -l

$ ./a.out -Wl,-le

- 61 -

-li

The -li option outputs diagnostic messages for all error levels.

-lw

The -lw option outputs diagnostic messages for error levels w, e, s, and u.

-le

The -le option outputs diagnostic messages for error levels e, s, and u.

-ls

The -ls option outputs diagnostic messages for error levels s and u.

-mu_no 0 <= u_no <= 2147483647

The -m option connects the specified unit number u_no to the standard error output file where diagnostic messages are to be written.If the -m option is omitted, unit number 0, the system default, is connected to the standard error output file. See Section "3.8 UsingEnvironment Variables for Execution", and Section "7.2 Unit Numbers and File Connection" for details.

Example: Runtime option -m

$ ./a.out -Wl,-m10

-n

The -n option controls whether prompt messages are sent to standard input. When the -n option is specified, prompt messages areoutput when data is to be entered from standard input using formatted sequential READ statements, including list-directed and nameliststatements. If the -n option is omitted, prompt messages are not generated when data is to be entered from standard input using aformatted sequential READ statement.

Example: Runtime option -n

$ ./a.out -Wl,-n

-pu_no 0 <= u_no <= 2147483647

The -p option connects the unit number u_no to the standard output file. If the -p option is omitted, unit number 6, the system default,is connected to the standard output file. See Section "3.8 Using Environment Variables for Execution", and Section "7.2 Unit Numbersand File Connection" for details.

Example: Runtime option -p

$ ./a.out -Wl,-p10

-q

The -q option specifies whether to capitalize the E, EN, ES, D, Q, G, L and Z edit output characters produced by formatted outputstatements. This option also specifies whether to capitalize the alphabetic characters in the character constants used by the inquiryspecifier (excluding the NAME specifier) in the INQUIRE statement. If the -q option is specified, the characters appear in uppercaseletters. If the -q option is omitted, the characters appear in lowercase letters. In Fortran95 and Fortran 2003, the characters appear inuppercase letters so the -q option is not required.

Example: Runtime option -q

$ ./a.out -Wl,-q

-ru_no 0 <= u_no <= 2147483647

The -r option connects the unit number u_no to the standard input file during execution. If the -r option is omitted, unit number 5, thesystem default, is connected to the standard input file. See Section "3.8 Using Environment Variables for Execution", and Section "7.2Unit Numbers and File Connection" for details.

Example: Runtime option -r

$ ./a.out -Wl,-r10

- 62 -

-tsec 1 <= sec <= 9999

The -t option sets the time limit in seconds for program execution. If -t20 is specified, for example, execution is canceled after 20seconds.

Example : Runtime option -t

$ ./a.out -Wl,-t20

-x

The -x option determines whether blanks in numeric edited input data are ignored or treated as zeros. If the -x option is specified,blanks are changed to zeros during numeric editing with formatted sequential input statements for which no OPEN statement has beenexecuted. The result is the same as when the BLANK= specifier in an OPEN statement is set to zero. If the -x option is omitted, blanksin the input field are treated as null and ignored. The result is the same as if the BLANK= specifier in an OPEN statement is set toNULL or if the BLANK= specifier is omitted.

Example: Runtime option -x

$ ./a.out -Wl,-x

-C or -Cu_no 0 <= u_no <= 2147483647

The -C option specifies how to process an unformatted file of IBM370-format floating-point data using an unformatted input/outputstatement. When the -C option is specified, the data of an unformatted file associated with the specified unit number is regarded asIBM370-format floating-point data in an unformatted input/output statement. The option argument u_no specifies an integer from 0to 147483647 as the unit number. If option argument u_no is omitted, the -C option is valid for all unit numbers connected to unformattedfiles. When the specified unit number is connected to a formatted file, the option is ignored for the file. When the -C option is notspecified, the data of an unformatted file associated with unit number u_no is regarded as IEEE-format floating-point data in anunformatted input-output statement.

Example: Runtime option -C

$ ./a.out -Wl,-C10

-G or -Gu_no 0 <= u_no <= 2147483647

The -G option specifies how to process an unformatted file of IBM370 EBCIDC character code data using an unformatted input/outputstatement. When the -G option is specified, the data of an unformatted file associated with the specified unit number is regarded asIBM370 EBCDIC character code data in an unformatted input/output statement. The option argument u_no specifies an integer from0 to 2147483647 as the unit number. If option argument u_no is omitted, the -G option is valid for all unit numbers connected tounformatted files. When the specified unit number is connected to a formatted file, the option is ignored for the file. When the -Goption is not specified, the data of an unformatted file associated with unit number u_no is regarded as ASCII character code data inan unformatted input-output statement.

Example: Runtime option -G

$ ./a.out -Wl,-G10

-Lb

Service routines use eight-byte logicals instead of four-byte logicals as arguments and function results.

Example: Runtime option -Lb

$ ./a.out -Wl,-Lb

-Li

Service routines use eight-byte integers instead of four-byte integers as arguments and function results.

Example: Runtime option -Li

$ ./a.out -Wl,-Li

-Lr

Service routines use eight-byte reals instead of four-byte reals as arguments and function results.

- 63 -

Example: Runtime option -Lr

$ ./a.out -Wl,-Lr

-Lu

If the length of the unformatted Fortran record to be transferred in a sequential access unformatted input/output statement is 2G bytesor more, input/output is by dividing it into multiple Fortran records. If the -Lu option is specified, the size of the field which exists inthe top and the end of the logical record and which is assigned the length of Fortran record is expanded to 8 bytes from 4 bytes, andcan be input/output as single Fortran record even if the length of the Fortran record is 2G bytes or more.

The data which has written to the -Lu option can be read only by unformatted sequential input statements with the -Lu option. Andthe data which has written without the -Lu option can be read only by that statements without the -Lu option.

If the -Lu option is specified differently across writing and reading the data, the unformatted sequential input statement may not work.

Example: Runtime option -Lu

$ ./a.out -Wl,-Lu

-M

The -M option specifies whether to output the diagnostic message (jwe0147i-w) when bits of the mantissa are lost during conversionof IBM370-IEEE-format floating-point data. If the -M option is specified, a diagnostic message is output if conversion of IBM370-IEEE-format floating-point data results in a bits of the mantissa being lost. When the -M option is omitted, the diagnostic message(jwe0147i-w) is not output.

Example: Runtime option -M

$ ./a.out -Wl,-M

-Q

The -Q option suppressed padding of an input field with blanks when a formatted input statement is used to read a Fortran record. Thisoption applies to cases where the field width needed in a formatted input statement is longer than the length of the Fortran record andthe file was not opened with and OPEN statement. The result is the same as if the PAD= specifier in an OPEN statement is set to NO.If the -Q option is omitted, the input record is padded with blanks. The result is the same as when the PAD= specifier in an OPENstatement is set to YES or when the PAD= specifier is omitted.

Example: Runtime option -Q

$ ./a.out -Wl,-Q

-Re

Disables the runtime error handler. The diagnostic message, traceback, error summaries, user control of errors by ERRSET andERRSAV, and execution of user code for error correction are suppressed. The standard correction is processed if an error occurs.

Example : Runtime option -Re

$ ./a.out -Wl,-Re

-Rp

The -Rp option issues a diagnostic message under certain circumstances for programs running on multiple processors.

If the -Rp option is specified, a diagnostic message (jwe1034i-w) is output when a program compiled with the compiler option -Kparallel runs on only one processor.

If the -Rp option is not specified, the diagnostic message (jwe1034i-w) is not output.

Example : Runtime option -Rp

$ ./a.out -Wl,-Rp

-Ry

A diagnostic message may be output when a program used service routines that returns the year using only the last two digits.

The last two digits of the year are returned when one of the following procedures is used.

- 64 -

- DATE service subroutine

- JDATE service function

If the return values are compared by a user program, a malfunction may occur. Thus the program should receive proper consideration.By specifying the runtime option -Ry,-li, the diagnostic message is output when these routines are executed.

The routines that return a four-digit number for the year are listed below:

- CTIME service function

- FDATE service subroutine

- IDATE service subroutine

- GETDAT service subroutine

- DATE_AND_TIME intrinsic subroutine

Example: Runtime option -Ry

$ ./a.out -Wl,-Ry,-li

-T or -Tu_no 0 <= u_no <= 2147483647

Little endian integer data, logical data, and IEEE floating-point data is transferred in an unformatted input/output statement.

The argument u_no is a unit number connected with an unformatted file. If u_no is omitted, -T takes effect for all unit numbers. Ifboth -T and -Tu_no are specified as in the example, "-Wl,-T,-T10", -T takes effect for all unit numbers.

Example: Runtime option -T

$ ./a.out -Wl,-T10

3.4 Execution Command Environment variableThe environment variable FORT90L may be used to specify runtime options. The runtime options specified in FORT90L are combinedwith the execution command option arguments. The execution command option arguments take precedence over the corresponding optionsspecified in the environment variable FORT90L.

The following examples show how to use the environment variable FORT90L.

Example 1:

a. When using the environment variable FORT90L

$ export FORT90L='-Wl,-e99,-le'$ ./a.out -Wl,-m99 -k

b. When not the using environment variable FORT90L

$ ./a.out -Wl,-e99,-le,-m99 -k

Both two above examples have the same meaning. When executing the command a.out, runtime options -e99, -le, and -m99, and user-defined executable program option -k are in effect.

Example 2:

$ export FORT90L='-Wl,-e10'$ ./a.out -Wl,-e99

When executing execution command a.out, runtime option -e99 is in effect.

3.5 Execution Profile FileThe default of runtime options can be changed by specifying the execution profile file (/etc/opt/FJSVmxlang/jwe_prof).

An execution profile file must be used according to the following format:

- 65 -

- Lines beginning with the pound sign (#) are treated as comment lines.

- A single runtime option cannot extend over multiple lines.

- The format of runtime options that can be specified must follow the conventions explained in Section 3.3 Runtime Options.

The following is an example of an execution profile file specification:

#Default options-Wl,-x,-le

The priority of runtime options is as follows, highest first:

1. Execution command operand

2. Environment variable

3. Profile file

4. Default value

3.6 Execution Command Return ValuesThe following table lists return values set by the execution command.

Return value Status

0 No error or level i (information message)

4 Level w error (warning)

8 Level e error (medium)

12 Level s error (serious)

16 Limit exceeded for level w, e, s error, or a level u error (unrecoverable) was detected

240 Abnormal termination

Others Forcible termination

3.7 Standard Input, Output, and Error Output for ExecutionThe default unit numbers for standard input, output, and error output for execution commands are:

Standard input : Unit number 5

Standard output : Unit number 6

Standard error output : Unit number 0

3.8 Using Environment Variables for ExecutionThis section describes environment variables that control execution.

fuxx filen

The fuxx environment variable connects units and files. The value xx is a unit number. The value filen is a filename to be connectedto a unit number. See Section "7.2 Unit Numbers and File Connection" for information on connecting units and files. The standardinput and output files (fu05 and fu06) and error file (fu00) must not be specified.

The following example shows how to connect data.file to unit number 10 prior to the start of execution.

$ export fu10="data.file"

- 66 -

fuxxbf size

The fuxxbf environment variable specifies the size of the input/output buffer used by a sequential or direct access input/output statementfor every unit number. The value xx in the fuxxbf environment variable specifies the unit number. The size argument used for sequentialaccess input/output statements is in kilobytes; the size argument used for direct access input/output statements is in Fortran records.The size argument must be an integer with a value of 1 to 2097151. The size argument for direct access input/output statements is thenumber of Fortran records per block in fixed-block format. The size argument must be an integer from 1 to 2147483747/RECL=specifier for an OPEN statement that indicates the number of Fortran records per block. If the value of size argument is incorrect, theinput/output buffer size is the default size. The default size is 8M bytes. If the environment variable and the runtime options -d or -gare specified at the same time, the environment variable is effective. However, if the size argument is not effective, the value specifiedwith the runtime options is effective.

When sequential access input/output statements are executed for unit number 10, the statements use an input/output buffer of 64kilobytes.

$ export fu10bf="64"

When direct access input/output statements are executed for unit number 10, the number of Fortran records included in one block is50. The input/output buffer size is obtained by multiplying 50 by the value specified in the RECL= specifier of the OPEN statement.

$ export fu10bf="50"

FLIB_ALLOC_ALIGN size

The boundary of allocated memory by an ALLOCATE statement is aligned on a specified alignment boundary size. The value of sizemust be 16 or more and must be a power of two.

The following example shows how to align 128 byte boundary to allocated memory.

$ export FLIB_ALLOC_ALIGN=128

FLIB_EXCEPT u

The environment variable FLIB_EXCEPT=u controls floating point underflow interrupt processing.

If the program was compiled with the -NRtrap compiler option and the program was executed setting the environment variableFLIB_EXCEPT=u, the system performs floating point underflow interrupt processing. The system outputs diagnostic messagejwe0012i-u during execution.

If the program was compiled without the -NRtrap compiler option or the program was executed without setting the environment variableFLIB_EXCEPT=u, the system ignores the floating point underflow.

The following example shows how to specify FLIB_EXCEPT.

$ export FLIB_EXCEPT=u

FLIB_HOOK_TIME time

Specifies the interval at which the user-defined subroutine is called at regular time interval. When time is specified in the environmentvariable FLIB_HOOK_TIME, the user-defined subroutine is called at interval of time milliseconds.

The unit for specifying time is milliseconds, and the valid value range is 0 <= time <= 2147483647. If 0 is specified for time, callingat regular time interval is disabled. If the environment variable is not specified, the Fortran system default value takes effect. Thedefault value is 60000 milliseconds (one minute).

Refer to Section "8.2.3 Hook Function", for information about the hook function.

An example of specifying the environment variable FLIB_HOOK_TIME is shown below.

$ export FLIB_HOOK_TIME=600000

In this example, the user-defined subroutine is called at 10 minutes interval.

FLIB_IOBUFCPY

FLIB_IOBUFCPY is used for I/O buffer parallel transfer function. Only "MP" can be specified to FLIB_IOBUFCPY. See Section"12.4 I/O Buffer Parallel Transfer" for details.

- 67 -

$ export FLIB_IOBUFCPY="MP"

FLIB_IOBUFCPY_SIZE size

The environment variable FLIB_IOBUFCPY_SIZE is used to change the condition to execute I/O buffer parallel transfer function.When the size is specified to the environment variable FLIB_IOBUFCPY_SIZE, an unformatted I/O statement transfers data in parallelif data transfer size and I/O buffer size are more than size Kbytes. The size argument is in Kilobytes. The size argument must be aninteger with a value of 1 or more.

The following example shows how to specify FLIB_IOBUFCPY_SIZE.

$ export FLIB_IOBUFCPY_SIZE="16"

See Section "12.4 I/O Buffer Parallel Transfer" for information on I/O buffer parallel transfer.

FLIB_UNDEF_VALUE hex

This environment variable change the values used for undefined data reference checks. Using hex, hexadecimal values can be specified.The valid range of values is 00 <= hex <= FF. If the environment variable is not specified or if hex value is disabled, the Fortran systemdefault value takes effect. The default value is 8B.

See Section "8.2.1.3 Checking References for Undefined Data Items (UNDEF)" for details on undefined data reference checks.

An example of specifying the FLIB_UNDEF_VALUE environment variable is shown below.

$ export FLIB_UNDEF_VALUE=8C

In this example, the value 8C is applied in each byte of the data value used in the undefined data reference check.

FLIB_USE_STOPCODE n

This environment variable specifies whether to reflect the stop-code of the STOP/ERROR STOP statement in the return code. Theleast significant byte (0 to 255) in a stop-code is reflected in the return code if the value of n is 1. The stop-code is not reflected in thereturn code if anything other than 1 is specified to n or this environment variable is not specified.

The following example shows how to specify the environment variable FLIB_USE_STOPCODE.

$ export FLIB_USE_STOPCODE=1

In this example, the stop-code of the STOP/ERROR STOP statement is reflected in the return code.

TMPDIR dir

The TMPDIR environment variable changes the directory where an unnamed file is created. If this is not effective, an unnamed fileis created on the current directory.

The following example shows how to specify TMPDIR.

$ export TMPDIR='/tmp'

In this example, an unnamed file is created on /tmp.

3.9 Notes for Execution

3.9.1 Notes for Execution of the Non-process Parallel Job on the FX100System

Blurring or deterioration in the execution performance might happen because the access time of the memory is not uniform with the NUMAnode where the memory is allocated in the NUMA architecture.

The problem can be evaded by fixing the NUMA node used by using the numactl command as follows for non-process parallel job whosenumber of threads is 16 (the number of all CPUs on one NUMA node) or less.

Example: Start of program using numactl command

$ numactl -C 0 ./a.out

- 68 -

The cpuid of NUMA node used is adjusted to 0 by the -C option.

- 69 -

Chapter 4 Messages and ListingsThis chapter describes the system messages and listing produced as a result of compiling a Fortran source program or executing a Fortranprogram.

4.1 Compilation Output

4.1.1 Compilation Diagnostic MessagesThe output from the frtpx command and the Fortran compiler includes diagnostic messages. If frtpx command line arguments containerrors or if errors or warnings are detected during the compilation of Fortran source programs, diagnostic messages are written.

Compiler Messages

Formats

Messages issued by frtpx are in the following format:

$ frtpx: Message-text

Messages issued by the Fortran compiler are in the following format:

jwdxxxxz-y filename line column Message-text

The elements of such a compiler message are defined as follows:

jwdxxxxz-y

jwd: A diagnostic message from the Fortran compiler

xxxx: Message serial number

z: Message type (i, o, s, or p)

Message type Explanation

i This message is unrelated to the optimization.

o This message reports the optimization status.

s This message reports the optimization of using SIMD instructions.

p This message reports the optimization status of automatic parallelization.

y: Message error level (i, w, s, or u) and return code

Errorlevel

Returncode

Explanation

i0

No error.

The message is information.

w The message is a warning. Processing continues.

s

1

A serious error.

The system ignores the erroneous statement and continues processing other statements.

uAn unrecoverable error.

The system terminates processing.

filename

The filename where the message event was detected

- 70 -

line

The line number where the message event was detected

column

The column number of the line where the message event was detected. The column number may be added to diagnostic messages.

Message-text

The message text is issued in either English or Japanese depending on user environment variables, such as LC_MESSAGES andLANG. Using compiler options, the user can specify either long or short message text.

4.1.2 Compilation Information

4.1.2.1 Compilation Information OptionsCompilation information is output by specifying the -Q, -Nlst or -Nlst_out=file compiler option. The compilation information is outputto a file with suffix .lst. If more than one Fortran source program file is specified, it is output to the first-file-name.lst. The compilationinformation can be controlled by specifying the arguments of -Q, -Nlst or -Nlst_out=file. See Section 4.1.2.2 Example of CompilationInformation. The options and their meanings are as follows:

-Q or -Nlst

Outputs the source program list and diagnostic error messages.

-Qa or -Nlst=a

Outputs attributes of names in addition to the compilation information specified by the -Q or -Nlst option.

-Qd or -Nlst=d

Outputs information about derived type layout in addition to the compilation information specified by the -Q or -Nlst option.

-Qi or -Nlst=i

Outputs include file listings and a list of the names of include files in addition to the compilation information specified by the -Q or -Nlst option.

-Qm or -Nlst=m

If the -Qm or -Nlst=m option is specified, the Fortran program which expresses the situation of automatic parallelization with OpenMPdirectives is output in addition to the compilation information produced by the -Qp or -Nlst=p option. "!FRT" is added at the last ofdirectives output by this function in order to distinguish from directives already described.

In addition, when the situation of automatic parallelization cannot be correctly expressed with OpenMP directive, the line which startsnot in "!$OMP" but in "!!!!! PARALLEL INFO:" is output. The situation of the optimization performed simultaneously with automaticparallelization may be added to the line.

1. The kind of OpenMP directive output by this function is shown below.

- PARALLEL DO

- FIRSTPRIVATE(var_name,...)

- LASTPRIVATE(var_name,...)

- PRIVATE(var_name,...)

- REDUCTION(+:var_name)

- REDUCTION(-:var_name)

- REDUCTION(*:var_name)

- REDUCTION(MAX:var_name)

- REDUCTION(MIN:var_name)

- REDUCTION(.OR.:var_name)

- REDUCTION(.AND.:var_name)

- 71 -

2. The kind of information output to the line which starts in "!!!!! PARALLEL INFO:" is shown below. Referring to these information,user can easily rewrite the program into OpenMP form.

- LOOP INTERCHANGEDIt means that loop was interchanged.

- LOOP FUSEDIt means that loop was fused.

- LOOP SPLITIt means that loop was split.

- COLLAPSEDIt means that nested do loops were collapsed into a single loop was performed.

- OTHER METHODSIt means that the situation of automatic parallelization cannot be correctly expressed with OpenMP directive by a reason otherthan the above. The example of the main factors is shown below.

- The loop is not DO loop.

- The loop including an induction variable.

- The loop including a variable of derived type.

- The loop including a definition in the array element with an invariable subscript.

- The variable of reduction operation is the array element with an invariable subscript.

- The initial value,terminal value or incrementation value of loop including the array element, function reference, and so on.

- UNKNOWN VARIABLE(var_name)It means that it is the variable that kind of OpenMP directive is not decided correctly. The example of the main factors is shownbelow.

- The variable that kind of OpenMP directive is not decided correctly because of optimization.

- The variable of complex type.

- Clause of OpenMP(var_name)When the kind of OpenMP directive is not decided correctly, the information for the clause of OpenMP might be output as areference information. The kind of clause is shown below:

- FIRSTPRIVATE(var_name,...)

- LASTPRIVATE(var_name,...)

- PRIVATE(var_name,...)

- REDUCTION(+:var_name)

- REDUCTION(-:var_name)

- REDUCTION(*:var_name)

- REDUCTION(MAX:var_name)

- REDUCTION(MIN:var_name)

- REDUCTION(.OR.:var_name)

- REDUCTION(.AND.:var_name)

3. Notes on use are shown below.

- When -Qm or -Nlst=m option is specified to file created by -P option, the file having name that .omp is added to the file namebefore preprocessing is created.

$ frtpx -P file.F95 # file.cpp.f95 was created.$ frtpx -Qm -Kparallel file.cpp.f95

- 72 -

# file.omp.F95 was created.$ frtpx -Nlst=m -Kparallel file.cpp.f95 # file.omp.F95 was created.

- If OpenMP directive REDUCTION([MAX|MIN]:var_name) is created from Fortran program containing a variable or a programunit with the name MAX or MIN, the Fortran program can issue the error of S level and cannot be compiled. Correct Fortranprogram not to use the name MAX or MIN in that case.

- Even when parallelized, neither OpenMP directive nor "!!!!! PARALLEL INFO:" line may be output by the influence ofoptimization.

-Qofile or -Nlst_out=file

Outputs the information to a file named 'file'.

-Qp or -Nlst=p

This option specifies optimization information and automatic parallelization information to be output along with listing of -Q or -Nlstoption. Optimization information shows status of parallelization, inlining and unrolling. To know the output on statements specifiedwith the OpenMP Fortran directives, see 12.3 Parallelization by OpenMP Specifications.

1. Automatic parallelization information

The meanings of marks for DO statement or array expression are as follows.

These describe the parallelization status of whole DO loop or array expression.

p : DO loop or array expression is parallelized

m : DO loop or array expression is partially parallelized

s : DO loop or array expression is not parallelized

blank : DO loop or array expression is not target of parallelization

When the above mark "p", "m" or "s" is displayed for DO statement, "p", "m" or "s" is displayed for executable statement in loop.The meaning of marks for executable statements is as follows.

p : Executable statements can be parallelized

m : Executable statements can be partially parallelized

s : Executable statements cannot be parallelized

The following condition applies to statements targeted for parallelization. Non-executable statements and executable statementsthat process only transfer of control are displayed by symbols determined by the parallelization information of executable statementsbefore and after these statements. Transfer of control statements include ELSE, END IF, CONTINUE, ENDDO, ELSEWHERE,and so on. Also parallelization information for these statements may be blank. Even when a DO loop or array expression is enabledfor parallelization, for the sake of performance, the system may not parallelize it. In this case it displays "s" for the DO statementand displays the unchanged analysis results for statements inside the DO loop.

2. Optimization information

The meaning of optimization marks are as follows.

i : inlined the reference to procedure

number : unrolled expansion number

f : full unrolling the loop

-Qt or -Nlst=t

If the -Qt or -Nlst=t option is specified, the details optimization information and the statistics information is output in addition to thecompilation produced by the -Qp or -Nlst=p option.

The optimization information is output in loop units. The parallel dimension information is output in statement units. The statisticsinformation is output approximate stack size is used in a reference of procedure.

- 73 -

1. The outputs information in loop units is defined as follows:

a. Displaying about automatic parallelization

- Standard iteration count : NWhen the iteration count of a loop is N or more, the loop is brought to parallel execution. When less than N, it is broughtto serial execution.

b. Displaying optimization information in loop units

- INTERCHANGED(nest:nest-no)It means that loop was interchanged. nest-no is the nesting level of the loop that was moved by loop interchange optimization.(nest:) may not be displayed if the other optimization like loop collapsing was applied.

- FUSED(lines:num1,num2[,num3]...)It means that loops were fused.Loops at line num2 and subsequent lines were fused into a loop at line num1.(lines:) isn't displayed at line num2 and subsequent lines, that were fused to line num1.

- SPLITIt means that loop was split.

- COLLAPSEDIt means that nested loops were collapsed into a single loop.

- SOFTWARE PIPELININGIt means that loop was performed software pipelining.

- SIMD(VL: length[, length]...)It means that SIMD instructions were generated for the loop.- A SIMD instruction treats length elements of array. When loop fission is applied to the loop and the values of lengthfor the each loop are different, two or more lengths are displayed.

- UXSIMDIt means that SIMD instructions were generated for the loop by UXSIMD.

- STRIPINGIt means that loop was performed striping.

- MULTI-OPERATION FUNCTIONIt means that loop includes multi-operation function.

- PATTERN MATCHING(matmul)It means that loop was changed library function call(matmul).

- UNSWITCHINGIt means that loop was performed loop unswitching.

- FULL UNROLLINGIt means that loop was fully unrolled.

- CLONEIt means that CLONE optimization was applied to the loop.

- LOOP VERSIONINGIt means that loop versioning was applied to the loop.

c. Information that relates to the prefetch instruction is displayed.

- PREFETCH: NIndicates the number of prefetch instructions that exist in the loop.

- variable name: N,...Indicates the number of prefetch instructions for the variable name.

- OTHER PREFETCHIndicates the number of prefetch instructions for the variable which was generated by compiler.

- 74 -

d. Information that relates to the xfill instruction is displayed.

- XFILL: NIndicates the number of xfill instructions that exist in the loop.

- variable name: N,...Indicates the number of xfill instructions for the variable name.

- OTHER XFILLIndicates the number of xfill instructions for the variable which was generated by compiler.

2. Marks for DO statement or array expression are as follows:

p : It means the beginning of parallelized range.(The mark "p" of -Qp or -Nlst=p option is output

together, so not p but "pp" is displayed)

3. The statistics information is defined as follows:

a. The statistics information is output approximate stack size is used in a reference of procedure.

Note:

- The stack size in internal procedure is displayed with stack size in parent procedure.

- The following data is allocated stack area, but the size does not include outputs in this function.- The data is specified THREADPRIVATE directive.- The data is automatic data object and -Kautoobjstack option in effect.

b. The number of prefetch instructions procedure is displayed.

4. SIMD information

The meanings of marks for DO statement or array expression are as follows.

These describe the SIMD optimization status of whole DO loop or array expression.

x : DO loop or array expression can be applied SIMD optimization by UXSIMD

v : DO loop or array expression is applied SIMD optimization

m : DO loop or array expression is applied partially SIMD optimization

s : DO loop or array expression is not applied SIMD optimization

blank : DO loop or array expression is not target of SIMD optimization

When the above mark "v", "m" or "s" is displayed for DO statement, "v", "m" or "s" is displayed for executable statement in loop.

The meaning of marks for executable statements is as follows.

x : Executable statements can be applied SIMD optimization by UXSIMD

v : Executable statements can be applied SIMD optimization (*1)

m : Executable statements can be applied partially SIMD optimization (*1)

s : Executable statements cannot be applied SIMD optimization

The following condition applies to statements targeted for SIMD optimization.

Non-executable statements and executable statements that process only transfer of control are displayed by symbols determined bythe SIMD optimization information of executable statements before and after these statements.

Transfer of control statements include ELSE, END IF, CONTINUE, ENDDO, ELSEWHERE, and so on.

Also SIMD optimization information for these statements may be blank.

Even when a DO loop or array expression is enabled for SIMD optimization, for the sake of performance, the system may not SIMDoptimization it. In this case it displays "s" for the DO statement and displays the unchanged analysis results for statements insidethe DO loop.

- 75 -

*1: When SIMD optimization is applied to a DO loop and the mark "v" or "m" is displayed for the DO statement, the mark "v"could be displayed for statements in the loop though SIMD extensions are not used for them. The mark "s" is displayed for a statement preventing SIMD optimization, which is a clue to tune programs.

-Qx or -Nlst=x

Outputs a cross reference listing of names and labels in addition to the compilation information specified by the -Q or -Nlst option.

These arguments can be specified at the same time, separated by comma as in -Qa,ofile,x or -Nlst=a,lst_out=file,lst=x.

4.1.2.2 Example of Compilation InformationThe following is an example of the output produced by the -Q or -Nlst option.

Main program "CHECK" (line-no.)(nest) 1 PROGRAM CHECK 2 INTEGER,DIMENSION(10,10) :: A = 0 3 INTERFACE 4 SUBROUTINE SUB(I,J,A) 5 INTEGER,DIMENSION(:,:) :: A 6 END SUBROUTINE 7 END INTERFACE 8 1 DO I=1,10 9 2 DO J=MOD(I,2)+1,10,2 10 2 CALL SUB(I,J,A) 11 2 END DO 12 1 END DO 13 WRITE (*,"(10I3)") (A(:,I),I=1,10) 14 END PROGRAM

Procedure information Lines : 14 Statements : 14

External subroutine subprogram "SUB" (line-no.)(nest) 15 16 SUBROUTINE SUB(I,J,A) 17 INTEGER,DIMENSION(:,:) :: A 18 A(I,J) = 1 19 END SUBROUTINE


Total information Procedures : 2 Total lines : 19 Total statements : 18

The following is an example of diagnostic error messages that are output.

Main program "ERROR" (line-no.)(nest) 1 PROGRAM ERROR 2 1 DO 10 I=1,3 3 2 DO 20 J=1,3 4 2 CALL SUB(I*J) 5 2 10 ENDDO 6 1 END PROGRAM

Diagnostic messages: program name(ERROR)

- 76 -

jwd1027i-s "error.f", line 3: Undefined statement label ' 20' referenced. jwd1131i-s "error.f", line 3, column 9: The program unit contains unmatched nest in DO construct, IF construct, CASE construct, SELECT TYPE construct, ASSOCIATE construct, CRITICAL construct, or BLOCK construct.


External subroutine subprogram "SUB" (line-no.)(nest) 7 8 SUBROUTINE SUB(I) 9 INTEGER,INTENT(IN) :: I 10 PRINT *,I() 11 END SUBROUTINE

Diagnostic messages: program name(SUB) jwd2008i-i "error.f", line 8: Dummy argument 'I' not used in this subprogram. jwd1771i-s "error.f", line 10, column 16: 'I' already has the INTENT attribute.



The following is an example from the -Qi or -Nlst=i option.

Module "MOD" (inc)(line-no.)(nest) 1 MODULE MOD 2 CHARACTER(LEN=15) CC 3 END MODULE


Main program "INCLUDE" (inc)(line-no.)(nest) 4 5 PROGRAM INCLUDE 6 INCLUDE "INC1" 1 1 use mod,only:cc 7 CALL INIT() 8 PRINT *,CC 9 END PROGRAM


External subroutine subprogram "INIT" (inc)(line-no.)(nest) 10 11 SUBROUTINE INIT() 12 INCLUDE "INC2" 2 1 include "inc1" 1 1 use mod,only:cc

- 77 -

2 2 13 CC = "FUJITSU FORTRAN" 14 END SUBROUTINE


Total information Procedures : 3 Total lines : 14 Total statements : 10 Include file name list 1 : ./inc1 2 : ./inc2

The following is an example of the output from the -Qp -Kparallel -x- or -Nlst=p -Kparallel -x- option.

Main program "MAIN" (line-no.)(nest)(optimize) 1 INTEGER,PARAMETER ::N=10000 2 REAL,DIMENSION(N)::A 3 1 p 6 DO I=1,N 4 1 pi 6 A(I) = A(I) * FUNC(I) 5 1 p 6 ENDDO 6 END


External function subprogram "FUNC" (line-no.)(nest)(optimize) 7 FUNCTION FUNC(I) 8 FUNC=I+1 9 END



The following is an example of the output from the -Qx or -Nlst=x option.

Module "COMPLEX" (line-no.)(nest) 1 MODULE COMPLEX 2 TYPE TYPE1 3 SEQUENCE 4 REAL :: R_PART 5 REAL :: I_PART 6 END TYPE TYPE1 7 INTERFACE OPERATOR(+) 8 MODULE PROCEDURE PLUS 9 END INTERFACE 10 PRIVATE PLUS 11 CONTAINS 12 TYPE(TYPE1) FUNCTION PLUS(OP1,OP2) 13 TYPE(TYPE1),INTENT(IN) :: OP1,OP2

- 78 -

|(Class and Type) : type name |(Attributes) : use-associated |(Declaration) : |(Definition) : |(Reference) : 21 [OPERATOR(+)] |(Class and Type) : user defined operator |(Attributes) : use-associated |(Declaration) : |(Definition) : |(Reference) : 23


The following is an example of the output from the -Qd or -Nlst=d option.

Module "COMPLEX" (line-no.)(nest) 1 MODULE COMPLEX 2 TYPE TYPE1 3 SEQUENCE 4 REAL :: R_PART 5 REAL :: I_PART 6 END TYPE TYPE1 7 INTERFACE OPERATOR(+) 8 MODULE PROCEDURE PLUS 9 END INTERFACE 10 PRIVATE PLUS 11 CONTAINS 12 TYPE(TYPE1) FUNCTION PLUS(OP1,OP2) 13 TYPE(TYPE1),INTENT(IN) :: OP1,OP2 14 PLUS%R_PART = OP1%R_PART + OP2%R_PART 15 PLUS%I_PART = OP1%I_PART + OP2%I_PART 16 END FUNCTION PLUS 17 END MODULE


Scoping unit of module : COMPLEX Derived type construction TYPE1 | (Attributes) : SEQUENCE |R_PART | (Type and Attributes) : REAL(4), PUBLIC |I_PART | (Type and Attributes) : REAL(4), PUBLIC

Scoping unit of module sub-program : PLUS

Main program "MAIN" (line-no.)(nest) 18 19 PROGRAM MAIN 20 USE COMPLEX 21 TYPE(TYPE1) CPX1, CPX2, RESULT 22 READ (*,*) CPX1, CPX2 23 RESULT = CPX1 + CPX2 24 PRINT *,RESULT

- 83 -

25 END PROGRAM MAIN


Scoping unit of program : MAIN


The following is an example of the output from the -Kparallel,reduction -Qm or -Kparallel,reduction -Nlst=m option.(.omp file )

SUBROUTINE SUB(N,M) REAL,DIMENSION(10000) :: A REAL,DIMENSION(10000,100) :: C,D REAL::RRES=0.0

CALL INIT(A,B,C)

!$OMP PARALLEL DO REDUCTION(MAX:RRES) !FRT DO I=1,N IF (RRES<A(I)) RRES = A(I) END DO CALL OUTPUT(RRES)!$OMP PARALLEL DO PRIVATE(I) !FRT DO J=1,M DO I=1,N C(I,J) = C(I,J) + D(I,J) END DO END DO CALL OUTPUT(C)

DO I=1,N!!!!! PARALLEL INFO: LOOP INTERCHANGED DO J=1,M D(I,J) = D(I,J) + C(I,J) END DO END DO CALL OUTPUT(D) END

The following is an example of the output produced by the -Kfast -Kparallel -Qt or -Kfast -Kparallel -Nlst=t option.

External subroutine subprogram "SUB" (line-no.)(nest)(optimize) 1 SUBROUTINE SUB(A,B,C,N) 2 IMPLICIT NONE 3 INTEGER(KIND=4) :: N 4 REAL(KIND=8),DIMENSION(N,N) :: A,B 5 REAL(KIND=8) :: C 6 INTEGER(KIND=4) :: I,J <<< Loop-information Start >>> <<< [OPTIMIZATION] <<< INTERCHANGED(nest: 2) <<< SIMD(VL: 4) <<< SOFTWARE PIPELINING <<< Loop-information End >>> 7 1 p 8v DO I=2,N <<< Loop-information Start >>> <<< [PARALLELIZATION] <<< Standard iteration count: 3

- 84 -

<<< [OPTIMIZATION] <<< INTERCHANGED(nest: 1) <<< Loop-information End >>> 8 2 pp 8 DO J=2,N 9 2 p 8v A(I,J) = A(I,J) + B(I,J) * C 10 2 p 8 END DO 11 1 p END DO 12 END SUBROUTINE

Procedure information Lines : 12 Statements : 12 Stack(byte): 32 Prefetch num: 0

Total information Procedures : 1 Total lines : 12 Total statements : 12 Total stack(byte): 32 Total prefetch num: 0

The following is an example of the output from the -Kparallel,uxsimd,eval -Qt or -Kparallel,uxsimd,eval -Nlst=t option.

External subroutine subprogram "SUB" (line-no.)(nest)(optimize) 1 SUBROUTINE SUB(NN,E,D) 2 IMPLICIT NONE 3 INTEGER,PARAMETER:: TN=8 ! THREAD NUM 4 INTEGER :: NN,K,I,J 5 REAL*8 :: D(8,NN,TN),E(8,NN,TN) 6 <<< Loop-information Start >>> <<< [PARALLELIZATION] <<< Standard iteration count: 2 <<< Loop-information End >>> 7 1 pp DO K=1,TN <<< Loop-information Start >>> <<< [OPTIMIZATION] <<< SOFTWARE PIPELINING <<< Loop-information End >>> 8 2 p DO I=1,NN <<< Loop-information Start >>> <<< [OPTIMIZATION] <<< SIMD(VL: 4) <<< UXSIMD <<< Loop-information End >>> 9 3 p 4vx DO J=1,8 10 3 p 4vx D(J,I,K) = 0.0000001D0*DBLE(I)/10000.D0 11 3 p 4vx E(J,I,K) = 0.0000002D0*DBLE(I)/10000.D0 12 3 p 4v ENDDO 13 2 p ENDDO 14 1 p ENDDO 15 END

4.1.2.3 Notes on Compilation InformationNotes on Compilation Information are shown in the following.

- 85 -

- The -Qp (-Nlst=p) option and -Qt (-Nlst=t) option may output wrong compilation information (optimization information) in thefollowing cases.

- When a function is inline expanded, different optimizations may be applied to each line. In this case, the compiler may outputinformation incorrectly as follows.

- Output messages redundantly.

- Output optimization messages inconsistent with compilation information (optimization information).

- Output no compilation information (optimization information).

In addition, PREFETCH: N, which is the number of prefetch instructions in the inlined function, includes all prefetch instructionsexpanded in multiple lines.

- The compiler applies multiple optimizations to one loop. In this case, the compiler may output information incorrectly as follows.

- Output optimization information on line number to an incorrect line.

- Output optimization messages inconsistent with compilation information (optimization information).

- When multiple loops are written in the same line in a program, the compiler outputs optimization information for only one of theloops. Write one loop in one line in order to output optimization information.

- Detailed optimization information for a loop which consists of IF constructs and GO TO statements is not output. Optimizationinformation on line number may be output to an incorrect line.

- The compiler may generate loops when any of the following conditions is met. The optimization messages may be output for thegenerated loops, and compilation information (optimization information) may be output for the generated loops when -Qp (-Nlst=p)or -Qt (-Nlst=t) option is specified.

- The actual argument is a pointer array, assumed shape array, or array section, and the corresponding dummy argument is anonpointer and nonassumed shape array.

- -Nquickdbg=undef, -Nquickdbg=undefnan, or -Nsetvalue=array option is effective.

- An array variable name appears in FIRSTPRIVATE, LASTPRIVATE, REDUCTION, COPYIN, or COPYPRIVATE clause ofOpenMP.

- There is a loop parallelized automatically by -Karray_private option or optimization control specifier ARRAY_PRIVATE,FIRST_PRIVATE or LAST_PRIVATE.

- When link time optimization is applied, the optimization which is different from the optimization displayed in compilation informationmay be applied at runtime.

- When a #line directive is included in the source program, compilation information (optimization information) output by -Qp (-Nlst=p)or -Qt (-Nlst=t) option is output based on the line number specified by the #line directive. Suppose, for example, a #line directivespecifying the line 3 in the source program is attached to a DO statement which is in the line 10. In this case, the compilation information(optimization information) about the DO statement is output to the line 3 in the source program.

4.2 Executable Program OutputProgram execution generates the following output:

- Print data from output statements

- Messages from PAUSE or STOP/ERROR STOP statements

- Diagnostic messages

- Trace back map

- Error summary information

- Debug output

This section describes the diagnostic messages, trace back map, and error summary information output when errors are detected duringFortran program execution. See 7.7 Data Transfer Input/output Statements for output statements. See 6.5.6 PAUSE Statement, and 6.5.7STOP/ERROR STOP Statement for PAUSE or STOP/ERROR STOP statements. See 8.2 Debugging Functions for debugging function.

- 86 -

4.2.1 Diagnostic MessagesA diagnostic message is output if PAUSE statements or STOP/ERROR STOP statements are executed or an error is detected during Fortranprogram execution.

Messages issued by the Fortran library are in the following format:

jwexxxxt-y Additional-information Message-text

a. jwexxxxt-yjwe : A diagnostic messages from the Fortran libraryxxxx : A message serial numbert : Output is 'i' or 'a'. 'i' is output for messages not requiring a response. 'a' is output by a PAUSE statement, which requires a response.See 6.5.6 PAUSE Statement for the PAUSE statement.

y Explanation

i Not an error. The message is information message.

w The message is a warning. Processing continues.

e A medium error. The system ignores the erroneous statement and continues processing. Up to 10erroneous statements can be detected before the system terminates processing. The error number forsystem terminates can be controlled -enum runtime option and runtime error processing library.

s A serious error. The system ignores the erroneous statement and terminates processing.

u An unrecoverable error. The system terminates processing.

b. Additional-information

Additional information may be added to diagnostic messages. The additional information indicates the line number of the Fortransource program where the error was detected.

c. Message-text

The message text is issued in English.

4.2.2 Trace Back MapA trace back map is generated as specified in the error control table. This information indicates how the main program calls the programin which the error was detected. See 8.1 Error Processing for information about the error control table. See 8.1.7 Exception HandlingProcessing, and 8.2.2 Debugging Programs for Abend for exception handling. See 2.2 Compiler Options for a compiler option -Nnoline.

The trace back map is generated in the following format after a diagnostic message:

error occurs at name1 line s1 loc wwwwwwww offset xxxxxxxxname1 at loc yyyyyyyy called from loc zzzzzzzz in name2 line s2name2 at loc yyyyyyyy called from loc zzzzzzzz in name3 line s2name3 at loc yyyyyyyy called from o.s

name1

Intrinsic function name or entry name of the program unit where the error was detected. If the name is unknown, '????????' is assignedto name1.

name2

Intrinsic function name or entry name of the program unit where the error was detected. If the name is unknown, '????????' is assignedto name2.

name3

Entry name of the main program; o.s indicates the control program. If the name is unknown, '????????' is assigned to name3.

s1

Line number of the statement where the error was detected. If the compiler option -Nnoline is specified, it does not appear.

- 87 -

s2

Line number of the statement where the program unit is called. If the compiler option -Nnoline is specified, it does not appear.

wwwwwwww

Absolute address (hexadecimal) of the statement where the error was detected.

xxxxxxxx

Relative address (hexadecimal) of the statement where the error was detected, starting at the beginning of the program unit; not outputfor intrinsic function errors; if this address is unknown, it does not appear.

yyyyyyyy

Absolute address (hexadecimal) of the beginning of the program unit; if this address is unknown, it is '0000000000000000'.

zzzzzzzz

Absolute address (hexadecimal) of the statement that called the program unit; if this address is unknown, it is '0000000000000000'.

name1, name2, and name3 are processing names of the procedure names. The processing names and corresponding procedure names areshown in the following table:

Kind of procedure Processing name

Main-program MAIN__

Internal-subprogram ofmain-program

Main-program-name.internal-subprogram-name_

External-subprogram External-subprogram-name_

Internal-subprogram ofexternal-subprogram

External-subprogram-name.internal-subprogram-name_

Module-subprogram Module-name.module-subprogram-name_

Internal-subprogram ofmodule-subprogram

Module-name.module-subprogram-name.internal-subprogram-name_

Note1 : Processing names are uniformly lowercase unless the compiler option -AU is specified.

Note2 : The information in shared objects may be not correct.

4.2.3 Error SummaryThe following error summary information is output when program execution is completed.

However, the system errors whose error identification number is 900 to 1000 are not output.

error summary (Fortran)error number error level error count jwennnni y zzz : : :total error count = xxx

nnnn

Message serial number

y

Error level y(y:i, w, e, s, u) detected

zzz

Error count

xxx

Total error count of diagnostic messages

- 88 -

Chapter 5 Data Types, Kinds, and RepresentationsThis section describes the Fortran data types and kinds, and describes the machine representations of these data.

5.1 Internal Data RepresentationThe following describes the internal representation for each data type.

5.1.1 Integer Type DataKind 1 (one-byte) integer type data is represented internally as an 8-bit two’s complement binary number with values ranging from -128to 127.

Kind 2 (two-byte) integer type data is represented internally as a 16-bit two’s complement binary number with values ranging from -32768to 32767.

Kind 4 (four-byte) integer type data is represented internally as a 32-bit two’s complement binary number with values ranging from-2147483648 to 2147483647.

Kind 8 (eight-byte) integer type data is represented internally as a 64-bit two’s complement binary number with values ranging from-9223372036854775808 to 9223372036854775807.

In two’s complement representation, the first bit of each integer is the sign. For a positive integer, the sign bit is zero; for a negativeinteger, the sign bit is 1. For example, the kind 1 representation of -1 is Z'FF'. For integer zero, all bits are set to zero.

5.1.2 Logical Type DataKind 1 (one-byte) logical type data is represented internally as an 8-bit binary number. For TRUE, the value is Z'01'. For FALSE, all bitsare zero.

Kind 2 (two-byte) logical type data is represented internally as a 16-bit binary number. For TRUE, the value is Z'0001'. For FALSE, allbits are zero.

Kind 4 (four-byte) logical type data is represented internally as a 32-bit binary number. For TRUE, the value is Z'00000001'. For FALSE,all bits are zero.

Kind 8 (eight-byte) logical type data is represented internally as a 64-bit binary number. For TRUE, the value is Z'0000000000000001'.For FALSE, all bits are zero.

5.1.3 Real and Complex Type DataThe internal representations of all kinds of real type data conform to IEEE specifications.

The internal representations of the real and imaginary parts of all kinds of complex type data are the same as the internal representationsof the corresponding kinds of real type data. A complex type value requires twice the storage of a real type value of the same kind.

The following describes in detail how floating-point data is stored in memory.

Kind 4 (single precision) real and complex

The figure below shows the bit composition for sign (s), exponent (e), and fraction (f).

Default real storage format:

The table below shows the bit patterns for numbers stored as default real values.

- 89 -

Table 5.1 Default real valuesCommon name Storage format Value

0 00000000 0. 0e+00

+ maximum normal number 7f7fffff 3.40282347e+38

- maximum normal number ff7fffff -3.40282347e+38

+ minimum normal number 00800000 1.17549435e-38

- minimum normal number 80800000 -1.17549435e-38

+ maximum denormalized number 007fffff 1.17549421e-38

- maximum denormalized number 807fffff -1.17759421e-38

+ minimum denormalized number 00000001 1.40129846e-45

- minimum denormalized number 80000001 -1.40129846e-45

+infinity 7f800000 Inf (Infinity)

-infinity ff800000 -Inf

+quiet_NaN 7fc00000-7fffffff NaN (Not a number)

-quiet_NaN ffc00000-ffffffff -NaN

+signaling_NaN 7f800001-7fbfffff NaN

-signaling_NaN ff800001-ffbfffff -NaN

Kind 8 (double precision) real and complex

The figure below shows the bit composition of sign (s), exponent (e), and fraction (f).

Double-precision real storage format:

The table below shows the bit patterns for numbers stored as double-precision real values.

Double-precision real values

Common name Storage format Value

0 00000000 00000000 0.0e+00

+ maximum normal number 7fefffff ffffffff 1.797693134862316d+308

- maximum normal number ffefffff ffffffff -1.797693134862316d+308

+ minimum normal number 00100000 00000000 2.225073858507201d-308

- minimum normal number 80100000 00000000 -2.225073858507021d-308

+ maximum denormalized number 000fffff ffffffff 2.225073858507201d-308

- maximum denormalized number 800fffff ffffffff -2.225073858507021d-308

+ minimum denormalized number 00000000 00000001 4.940656458412465d-324

- minimum denormalized number 80000000 00000001 -4.940656458412465d-324

+infinity 7ff00000 00000000 Inf

-infinity fff00000 00000000 -Inf

+quiet_NaN 7ff80000 00000000 NaN

- 90 -


- 7fffffff ffffffff

-quiet_NaN fff80000 00000000

- ffffffff ffffffff

-NaN

+signaling_NaN 7ff00000 00000001

- 7ff7ffff ffffffff

NaN

-signaling_NaN fff00000 00000001

- fff7ffff ffffffff

-NaN

Kind 16 (quadruple precision) real and complex

The figure below shows the bit composition of sign (s), exponent (e), and fraction (f).

Quadruple-precision real storage format:

The table below shows the bit patterns for numbers stored as quadruple-precision real values.

Quadruple-precision real values


0 00000000 00000000 00000000 00000000 0.0e+00

+ maximum normalnumber

7ffeffff ffffffff ffffffff ffffffff 1.1897314953572317650857593266280070q+4932

- maximum normalnumber

fffeffff ffffffff ffffffff ffffffff -1.1897314953572317650857593266280070q+4932

+ minimum normalnumber

00010000 00000000 00000000 00000000 3.3621031431120935062626778173217526q-4932

- minimum normalnumber

80010000 00000000 00000000 00000000 -3.3621031431120935062626778173217526q-4932

+ maximumdenormalized number

0000ffff ffffffff ffffffff ffffffff 3.3621031431120935062626778173217519q-4932

- maximumdenormalized number

8000ffff ffffffff ffffffff ffffffff -3.3621031431120935062626778173217519q-4932

+ minimumdenormalized number

00000000 00000000 00000000 00000001 6.4751751194380251109244389582276465q-4966

- minimumdenormalized number

80000000 00000000 00000000 00000001 -6.4751751194380251109244389582276465q-4966

+infinity 7fff0000 00000000 00000000 00000000 Inf

-infinity ffff0000 00000000 00000000 00000000 -Inf

+quiet_NaN 7fff8000 00000000 00000000 00000000

-7fffffff ffffffff ffffffff ffffffff

NaN

-quiet_NaN ffff8000 00000000 00000000 00000000

- ffffffff ffffffff ffffffff ffffffff

-NaN

- 91 -


+signaling_NaN 7fff0000 00000000 00000000 00000001

-7fff7fff ffffffff ffffffff ffffffff

NaN

-signaling_NaN ffff0000 00000000 00000000 00000001

- ffff7fff ffffffff ffffffff ffffffff

-NaN

5.1.4 Character Type DataCharacter data is represented internally using ASCII codes. Each character is represented as an 8-bit binary number without a parity bit.

All ASCII characters, both printable and nonprintable, are allowed in character constants and comments.

5.1.5 Derived Type DataA derived type is a data type derived from the intrinsic data types, whose ultimate components are of intrinsic type. The derived type scalardata, which is a set of intrinsic type data, is called a structure.

If a derived type is not a sequence type, the storage sequence of the structure is undefined.

5.2 Correct Data BoundariesData must be aligned on the correct boundaries. The compiler usually aligns the data on the correct boundaries.

For elements of derived type and variable in a common block, the compiler assigns the data to the correct boundaries. However, if thecompiler option -AA is specified, the compiler may not align data on the correct boundaries. In this case, the boundaries are those specifiedby the user for variables that share storage space because of an EQUIVALENCE statement. For a derived type definition, the boundarythat follows the last component must be a correct boundary for the derived type.

For example, consider the following derived type definition when the compiler option -AA is specified:

TYPE TAG REAL(8) R INTEGER(4) IEND TYPE

The correct boundary for this derived type (TAG) is an address that is a multiple of 8. But because the boundary for the component I is amultiple of 4, but not 8, this derived type definition is incorrect.

Table 5.1 lists the correct data boundaries.

Table 5.2 Correct boundaries for data types

Data type Correct boundary

One-byte integer type Arbitrary address

Two-byte integer type Address in multiples of 2

Four-byte integer type Address in multiples of 4

Eight-byte integer type Address in multiples of 8

One-byte logical type Arbitrary address

Two-byte logical type Address in multiples of 2

Four-byte logical type Address in multiples of 4

Eight-byte logical type Address in multiples of 8

Single-precision real type Address in multiples of 4

Double-precision real type Address in multiples of 8

Quadruple-precision real type Address in multiples of 8

- 92 -

Data type Correct boundary

Single-precision complex type Address in multiples of 4

Double-precision complex type Address in multiples of 8

Quadruple-precision complex type Address in multiples of 8

Character type Arbitrary address

Derived type Address in multiples of the largest required

multiple for all of the components

Boundary of real type array

The first element of a single-precision real type array is allocated on 8-byte boundary, and double-precision real type array's is allocatedon 16-byte boundary. This is because these arrays may be applied SIMD optimizations.

However, arrays passed as arguments, arrays in a COMMON block which are not first element of the block and pointer arrays may notbe allocated on above boundary.

5.3 Precision ConversionThe precision of the arithmetic data that Fortran handles is restricted by the hardware. This section explains how to use the precision-improving and precision-lowering functions that the system provides.

5.3.1 Precision ImprovingThe precision-improving function increases the precision in programs. This function converts constants, variables, and functions of thespecified type to the next higher precision.

5.3.1.1 Precision-Raising OptionsPrecision improving can be executed for real and complex constants, variables, and functions.

If compiler option -Ad is specified, the data types are converted as follows:

Single-precision real

Single-precision complex

->

->

Double-precision real

Double-precision complex

If compiler option -Aq is specified, the data types are converted as follows:



->

->

Quadruple-precision real

Quadruple-precision complex

An example of precision improving for constants, variables, and functions when compiler option -Ad is specified as follows.

Example: Precision improving for constants, variables, and functions

IMPLICIT REAL(8)(D)REAL KNOTE,KINCOMPLEX ARY(10),CDATAKNOTE = 1.0KIN = 3.111-2.5D0*KNOTEDEX = 0.5D0-1.111111111111ARY(1) = (1.1,1.0)KIN = KIN-SQRT(KIN/2.0)

The above program is equivalent to the following program.

- 93 -

IMPLICIT REAL(8)(D)REAL(8) KNOTE,KINCOMPLEX(8) ARY(10),CDATAKNOTE =1.0_8KIN =3.111_8-2.5D0*KNOTEDEX =0.5D0-1.111111111111_8ARY(1) =(1.1_8,1.0_8)KIN =KIN-DSQRT(KIN/2.0_8)

If precision is improved for an intrinsic function, the type of the function result is changed to the higher precision type. The function nameis not changed if an EXTERNAL statement is used to declare a user-defined function.

For user-defined external functions, precision is improved for the function result. If the precision is improved for a function reference,functions must also be recompiled using the precision-improving function.

For intrinsic functions, precision improving affects function names, the types of arguments, and function results. Intrinsic functions withspecific names are changed to the appropriate name. If an intrinsic function has a generic name, the precision-improving function isexecuted for the arguments, and an appropriate version of the function is selected. If an intrinsic function is used as an actual argument,the name is changed to reflect precision improving.

5.3.2 Precision Lowering and Analysis of ErrorsThe precision-lowering function determines the effect that a rounding error has on the result. Rounding

errors occur because of the limit on floating-point data precision.

Numerical programs can be debugged using the precision-lowering function to check whether algorithm

errors are tolerable.

For example, consider programs that solve simultaneous linear equations or obtain eigenvalues or eigenvectors of matrices. In theseprograms, a slight input error in the matrix may cause an extremely large error. Thus, the user may not be able to determine whether theresult is correct. The precision-lowering function is provided to improve and debug numerical programs by detecting sections easilyaffected by errors. This facility helps the user select the appropriate solution to resolve the errors and minimize the effect of input errors.

Generally, the following relationship can be determined for the number of arithmetic digits and the precision of the answer:

d = (p - a)/m

d : Precision of answer (digits)p : Number of arithmetic digitsa : Missing digitsm : Values for the problem and solution method

The value of m is usually 1 and corresponds to the order of multiple roots when roots of algebraic equations are obtained. The precedingequation expresses the problem conditions precisely. However, the equation does not apply when:

- Input data or a constant is erroneous.

- A problem equation contains an error.

- A calculation terminates inappropriately.

That is, the above expression comes into existence when rounding error a prime cause. Incrementing and decrementing the digits by oneincrements and decrements the solution precision by 1/m digits. Thus, a slight shifting of the number of digits shifts the solution precisionby the same amount. The precision-lowering function uses this principle to obtain the solution precision.

Compiler option -Cl enables the precision-lowering function.

5.3.2.1 Defining the Precision-Lowering FunctionSpecifying the following option enables the precision-lowering function:

-Cl 0 ≤ l ≤ 15

The precision-lowering function can be used on all kinds (precisions) of real and complex data.

- 94 -

If the precision-lowering function is effect, when an assignment statement defines the all kinds of real and complex data, the low-order lbits are set to zero.

The precision-lowering number must be between 0 and 15, inclusive. The result when l is zero is equivalent to specifying compiler option-Cl. The PRNSET service subroutine can be called to change l dynamically. Thus, the precision-lowering function operates when l ischanged dynamically to a value other than zero.

The precision-lowering number l is determined for each executable program. Therefore, when the PRNSET service subroutine is calledfrom parallel region, the dynamic change by a specific thread may affect the execution of the another threads. The initial value of l isspecified when the main program is compiled. Compiler option -Cl only starts the precision-lowering function in subprograms; l itself ismeaningless in this case.

5.3.2.2 Using the Precision-Lowering FunctionIn floating-point calculations, l can be changed, the same calculation re-executed, and the results compared.

First time:

The program is compiled and executed using -C0.

Second time:

The program is compiled and executed using -C1.

If the digits from two consecutive attempts are identical, don't assume that either result is correct. If the digits from two consecutiveattempts are not identical, assume that one or the other of the two results is incorrect.

The number of executions can be increased to improve the accuracy. If you change the value of l to a number from 2 to 10 and increasethe number of executions, the precision of the result can be determined accurately. Generally, two executions are sufficient.

However, it is more efficient to use the computer for only one execution to determine the precision. Source code to determine the precisionfrom only one execution must be created. This program must use compiler option -Cl and include the PRNSET service subroutine. Usethe following procedure to create this source code:

1. Save the data to be changed in the check program section.

2. Execute CALL PRNSET(l).

3. Use the saved data in the calculations. Store the result in l.

4. Change l. Return to step 2 if the specified count is not reached.

5. Compare the results for each l. Print the information that matches.

An example of a program that uses the precision-lowering function as follows.

Example: Program that uses the precision-lowering function

REAL,DIMENSION(10,11)::A,B REAL,DIMENSION(10,10)::X 1 READ (5,FMT='(I2,I2)',END=99) N,M ! Enter data PRINT '(I2,I2)',N,M ! Print data DO I=1,N READ (5,'(5F12.8)') (A(I,J),J=1,N+1) WRITE(6,'(5F12.8)') (A(I,J),J=1,N+1) END DO DO L=1,M CALL PRNSET(L-1) CALL SWEEP(A,B,N) ! Calculate while changing! precision lowering X(:N,L)=B(:N,N+1) END DO WRITE(UNIT=*,FMT='(//,A,5X,3(A,10X),A)') & & ' -Cl ','X1=1.0','X2=2.0','X3=1.0','X4=-1.0' DO L=1,M WRITE(6,'(A,I1,A,1P8E16.7)') & & ' -C',L-1,' ',X(:N,L) ! Print the result END DO

- 95 -

WRITE(6,'(/,A,7X,3(A,14X),A)') ' -Cl:-C0','X1','X2','X3','X4' DO L=2,M X(:N,L)=X(:N,L)-X(:N,1) END DO DO L=2,M ! Result of -Cl and -C0 WRITE(6,'(A,I1,A,1P8E16.7)') ' -C',L-1,':-C0',X(:N,L) END DO GO TO 1 99 END SUBROUTINE SWEEP(A,Z,N) REAL,DIMENSION(10,11)::A,Z Z(:N,:N+1)=A(:N,:N+1) DO K=1,N Z(K,K+1:N+1) = Z(K,K+1:N+1)/Z(K,K) DO I=1,N ! Use the sweep method to solve the! simultaneous linear equations IF(K.EQ.I) CYCLE Z(I,K+1:N+1)=Z(I,K+1:N+1)-Z(I,K)*Z(K,K+1:N+1) END DO END DO END SUBROUTINE SWEEP

In the printed input data, the first line indicates the original values and test number. The second to fifth lines indicate the four coefficientsand one constant.

4 93.20000005 1.25979996 -2.01999998 5.13980007 -1.440199971.02010000 -1.35010004 3.09999990 -2.12010002 3.53999996-2.02979994 2.54999995 -1.37020004 3.64000010 -1.940000063.21000004 1.11020005 2.80999994 4.53999996 3.70040011

-Cl X1=1.0 X2=2.0 X3=1.0 X4=-1.0 -C0 9.9909294E-01 1.9976463E+00 1.0000743E+00 -9.9882913E-01 -C1 9.9581611E-01 1.9891458E+00 1.0003431E+00 -9.9459982E-01 -C2 9.9198890E-01 1.9792151E+00 1.0006566E+00 -9.8966026E-01 -C3 1.0101604E+00 2.0263634E+00 9.9916744E-01 -1.0131168E+00 -C4 1.0316334E+00 2.0820694E+00 9.9740791E-01 -1.0408325E+00 -C5 1.0174370E+00 2.0452271E+00 9.9856758E-01 -1.0225105E+00 -C6 8.9227295E-01 1.7205429E+00 1.0088348E+00 -8.6096954E-01 -C7 1.0482483E+00 2.1251831E+00 9.9604034E-01 -1.0623016E+00 -C8 9.1392517E-01 1.7766724E+00 1.0070496E+00 -8.8891602E-01

Difference between -Cl and -C0

-Cl:-C0 X1 X2 X3 X4 -C1:-C0 -3.2768250E-03 -8.5005760E-03 2.6881695E-04 4.2293072E-03 -C2:-C0 -7.1040392E-03 -1.8431187E-02 5.8233738E-04 9.1688633E-03 -C3:-C0 1.1067390E-02 2.8717041E-02 -9.0682507E-04 -1.4287710E-02 -C4:-C0 3.2540321E-02 8.4423065E-02 -2.6663542E-03 -4.2002678E-02 -C5:-C0 1.8343925E-02 4.7580719E-02 -1.5066862E-03 -2.3681164E-02 -C6:-C0 -1.0681915E-01 -2.7709961E-01 8.7604523E-03 1.3785934E-01 -C7:-C0 4.9155235E-02 1.2753677E-01 -4.0339231E-03 -6.3470840E-02 -C8:-C0 -8.5166931E-02 -2.2097397E-01 6.9752932E-03 1.0991287E-01

Notes:

1. This problem uses the sweep method to solve simultaneous linear equations that have poor characteristics (conditions).

2. When the number of precision-lowering bits varies from 0 to 8, this program prints the difference between the solution and the resultof -C0.

The difference in the result of using -C0 almost indicates an error, which shows that debugging is sufficient.

- 96 -

Chapter 6 Notes on ProgrammingThis chapter describes some of notes on programming on the Fujitsu Fortran system.

Input/output features are discussed in "Chapter 7 Input/output Processing".

6.1 Fortran VersionsFour versions of Fortran, all compiled using the frtpx command, are available:

1. FORTRAN66: Programs written in this version are called FORTRAN66 programs. Use compiler option -X6.

2. FORTRAN77: Programs written in this version are called FORTRAN77 programs. The compiler treats programs as FORTRAN77programs under the following conditions:

- The compiler option -X7 is specified.

- The file suffix is ".f", ".for", ".F" or ".FOR", and the compiler option -Xd7 is specified.

Example:

The compiler treats the following program as a FORTRAN77 program:

$ frtpx a.f90 -X7

$ frtpx b.f -Xd7

3. Fortran 95 (including Fortran 90): Programs written in this version are called Fortran 95 programs. The compiler treats programsas Fortran 95 programs under the following conditions:


- The file suffix is ".f90", ".f95", ".F90" or ".F95", and the compiler option -X is not specified.

Example:

The compiler treats the following program as a Fortran 95 program:

$ frtpx b.f -X9

$ frtpx a.f90

$ frtpx a.f95

4. Fortran 2003: Programs written in this version are called Fortran 2003 programs. The compiler treats programs as Fortran 2003programs under the following conditions:


- The file suffix is ".f03" or ".F03", and the compiler option -X is not specified.

- The file suffix is ".f", ".for", ".F" or ".FOR", and the compiler option -Xd7 is not specified.

Example:

$ frtpx b.f

$ frtpx a.f90 -X03

The compiler treats this program as a Fortran 2003 program:

Each Fortran Version, some items have different interpretations.

They are interpreted in accordance with any of the following Fortran version, and run.

- The Fortran version when the main program is compiled.

- The Fortran version of a program that includes a specification that is fixed when compiling a program

- 97 -

This compiler compiles all FORTRAN66, FORTRAN77, Fortran 95 (including Fortran 90) and Fortran 2003 programs. However,differences in the four versions can cause incorrect results.

The following items are interpreted differently, depending on whether compiler option -X6, -X7, -X9 or -X03 is specified.

The items and their interpretations are shown in the following.

These items are interpreted differently by language

specification.The interpretation is

decided when the eachprogram is compiled

The interpretation isdecided when the mainprogram is compiled

EXTERNAL statement that specifies an intrinsic function (Referto Section "6.6.2 An Intrinsic Function that Specifies EXTERNALAttribute")

*

DO loop iteration count method (Refer to Section "6.5.4.1 DOLoop Iteration Count Method")

*

Effect of the X edit descriptor (Refer to Section "7.17.5 Effect ofthe X Edit Descriptor")

*

REAL and CMPLX used as generic names (Refer to Section "6.6.3REAL and CMPLX Used as Generic Names")

*

Intrinsic procedures (Refer to Section "6.6.4 IntrinsicProcedures")

*

Intrinsic function names with the type declared (Refer to Section"6.6.5 Intrinsic Function Names with the Type Declared")

*

Assignment statements with overlapping character positions(Refer to Section "6.4.1 Assignment Statements with OverlappingCharacter Positions")

*

allocatable assignment (Refer to Section "6.4.2 AllocatableAssignment". )

*

SAVE attribute for initialized variables (Refer to Section "6.2.5SAVE Attribute for Initialized Variables")

*

OPEN statement with RECL= and with no ACCESS= specifier(Refer to Section "7.4.4 Default Value of ACCESS= Specifier withRECL=Specifier")

* *

INQUIRE statement (Refer to Section "7.6.2 Difference inLanguage Version of the INQUIRE statement")

*

Generalized integer editing (Refer to Section "7.17.1 GeneralizedInteger Editing")

*

Output character in formatted output statement (Refer to Section"7.17.2 Exponent Character in Formatted Output")

*

Result of G editing (Refer to Section "7.17.3 Result of G Editing") *

NAMELIST input/output (Refer to Section "7.7.4.3 NAMELISTInput/output Version Differences")

*

Diagnostic messages for a character string edit descriptor (Referto Section "7.17.4 Diagnostic Messages for the Character StringEdit Descriptor")

*

The form of nondelimited character constants produced by list-directed output (Refer to Section "7.7.5.2.1 The Form ofNondelimited Character Constants Produced by List-DirectedOutputt")

*

The specified value of the OPEN statement (Refer to Section "7.4.2Executing an OPEN Statement on a Connected File")

*

- 98 -

These items are interpreted differently by languagespecification.

The interpretation isdecided when the eachprogram is compiled

The interpretation isdecided when the mainprogram is compiled

Input/Output of Fortran records (Refer to Section "7.4.3 Input/output of Fortran Records")

*

Data transfer in direct access I/O (Refer to Section "7.7.3.3 TheError of Data Transfer in a DIRECT Access Input/output")

*

Input result of L editing (Refer to Section "7.17.6 L EditDescriptor")

*

Variable enclosed by parenthesis (Refer to Section "6.3.8 Variableenclosed by parenthesis")

*

The form of simple output list enclosed in parentheses (Refer toSection "6.5.9 Simple Output List Enclosed in Parentheses")

*

The output statement that starts by TYPE keyword (Refer toSection "6.5.10 The Output Statement that starts by TYPEKeyword")

*

List-Directed Input/Output Statements (Refer to Section "7.7.5.3List-Directed Input/output Version Differences")

*

The returned values of ATAN2, ATAN2D and ATAN2Q whennegative zero is specified for the argument (Refer to Section "6.6.7The returned values of the intrinsic function")

*

The returned values of LOG and SQRT when negative zero isspecified for imaginary part of the argument of complex type(Refer to Section "6.6.7 The returned values of the intrinsicfunction")

*

6.2 DataThis section discusses features related to data.

6.2.1 COMMON StatementWhen using a COMMON statement, note the following items:

6.2.1.1 Boundaries of Variables in Common BlocksThe variables in common blocks must be aligned on the correct boundaries. For more information, see Section "5.2 Correct DataBoundaries". If data is not aligned on the correct boundaries, the compiler inserts space in the storage area and aligns the data correctly.However, there are exceptions, as described below.

- If the compiler option -AA is specified, the compiler does not insert space.

- Variables of common entities with BIND attribute have interoperable boundary.

- When double precision or quadruple precision real or complex variables are not aligned on addresses that are multiples of eight, thesevariables are aligned on addresses that are multiples of four. But if the compiler option -Kdalign is specified, they are aligned onaddresses that are multiples of eight.

The beginning address of a common block is always a multiple of eight.

6.2.1.2 Common Blocks that Have Initial ValuesWhen either of the following is specified, an error is detected at linking:

- 99 -

1. In different files, common blocks with the same name or blank commons are initialized more than once.

Example 1:

Error 1 at linking:

! file name : a.f ! file name : b.fPROGRAM MAIN SUBROUTINE SUBCOMMON /COM/A,B COMMON /COM/A,BDATA A/1.0/ DATA B/2.0/CALL SUB PRINT *,A,BEND END

2. In different files, common blocks with the same name or blank commons that have different sizes are used, and the common blockthat does not have the maximum size of the same common blocks is initialized.

Example 2:

Error 2 at linking:

! file name : a.f ! file name : b.fPROGRAM MAIN BLOCK DATA BLKCOMMON /COM/A,B COMMON /COM/APRINT *,A DATA A/1.0/END END

6.2.2 EQUIVALENCE StatementEach equivalence object must be aligned on the correct boundary.

6.2.3 InitializationThis section describes items related to initialization.

6.2.3.1 Character InitializationIf a character constant is specified for the corresponding noncharacter type variable, the system outputs a level w diagnostic message. Ifa character string that exceeds the length of an element is specified as an initial value, the excess is not used to initialize the subsequentelement.

Example1: Initializing array names

CHARACTER(3) A(2)DATA A/'ABCDEF'/

Initial values in this case:

A(1): 'ABC'A(2): No initial value (undefined)

Example2: Initializing array element

CHARACTER(LEN=2) B(4)DATA B(2)/'ABCDE'/

Initial values in this case:

B(1) No initial value (undefined)B(2) 'AB'B(3) No initial value (undefined)B(4) No initial value (undefined)

The following examples give 'CD' and 'E' as initial values to B(3) and B(4).

- 100 -

CHARACTER(LEN=2) B(4)DATA B(2),B(3),B(4)/'AB','CD','E'/

or

CHARACTER(LEN=2) B(4)CHARACTER(LEN=6) BBEQUIVALENCE (B(2),BB)DATA BB/'ABCDE'/

6.2.3.2 Initialization with Hexadecimal, Octal, and Binary ConstantsTo initialize an array, the number of hexadecimal, octal, or binary constants must match the number of array elements. If these numbersdo not match, the remaining part is undefined. Thus, one hexadecimal, octal, or binary constant corresponds to one array element. If aconstant exceeds an element in size, the excess is truncated.

Example: Initializing array elements with hexadecimal values

INTEGER A(3)DATA A/Z'F0F0A100F2F2A202'/

Initial value of A(1) : F2F2A202

Initial value of A(2) : No initial value (undefined)

Initial value of A(3) : No initial value (undefined)

If a hexadecimal, octal, or binary constant that exceeds the size of an array element is specified, the high-order excess is truncated.

If a hexadecimal, octal, or binary constant is shorter than an array element, the space to the left of the constant is padded with zeroes.

6.2.3.3 Overlapping InitializationIf multiple initial values are specified for the same variable, the system outputs a level w diagnostic message and uses the last valuespecified in the source code. However, the code should be corrected.

6.2.4 Using CRAY PointersWhen using CRAY pointers, note the following items:

a. Do not save the address of a dummy argument in a subprogram.

Example: Incorrect use of a pointer (1)

SUBROUTINE SUB(A)POINTER (IP,IPP)INTEGER A(10),IPP(10)SAVE IP:IP = LOC(A(2)):END

b. Do not return the address of an actual argument as a result.


FUNCTION ADDR(I)INTEGER ADDR,I(5)ADDR = LOC(I(2))ENDPOINTER (IP,POINTER)INTEGER POINTER,DATA(5),ADDRIP = ADDR(DATA)

- 101 -

POINTER = 1END

c. The data referenced by a pointer must not overwrite other data, except when the address is explicitly defined in a program unit bythe intrinsic function LOC. Do not use the address in an assignment statement for which the value is copied.


POINTER (IP1,ARRAY1)POINTER (IP2,ARRAY2)INTEGER DATA(5),ARRAY1(5),ARRAY2(5)IP1 = LOC(DATA)ARRAY1 = 1IP2 = IP1ARRAY2 = 2 ! cannot use the address definedEND ! in assignment statement

6.2.5 SAVE Attribute for Initialized VariablesIn Fortran 95 and Fortran 2003, initialized variables have the SAVE attributes.

In FORTRAN66 and FORTRAN77, initialized variables do not have the SAVE attribute unless they are given the SAVE attributeexplicitly.

Example: The SAVE attribute for a variable with an initial value

SUBROUTINE SINTEGER::I=1I=I+1END

In Fortran 95, the previous program is the same as the following:

SUBROUTINE SINTEGER::I=1SAVE II=I+1END

6.2.6 Operation of Real or Complex type for Initialization ExpressionWhen real or complex type is operated in initialization expression of nonexecutable statement, the result may have some difference withinthe margin of error even if the same operation appears in executable statement.

6.3 ExpressionsThis section discusses features related to expressions and operations.

6.3.1 Integer Data OperationsIf integer data is assigned, the size is not checked. If the result of integer operations exceeds the range allowed, the result is undefined.This rule is especially critical for integer comparisons.

Example:

IF(I-J)1,2,3

Note:

In this example, the result of I-J can overflow. Change the expression to a comparison using relational operators in a logical IF statement.

If overflow occurs when evaluating the constant expression that defines a parameter, the value of the named constant is undefined.

- 102 -

6.3.2 Real Data OperationsReal data of any precision is processed as accurately as possible. However, machine language does not always represent data exactly.

Example: Decisions with no tolerance are allowed, but the result may not be what is expected

IF (A-0.1) 1,2,13 IF (B.EQ.C) GO TO 4

Note:

Even if 0.1 is the expected result of calculating A, processing does not always go to the statement with label 2. Even if B and C areexpected to be equal, processing does not always go to the statement with label 4.

Do not execute equality or inequality value comparisons with relational operations using real or complex data. Level i diagnostic messagesare produced for relational expressions that do not allow tolerance if the -Ec compiler option is specified. Relational expressions with notolerance allowed are "e1 relop e2", where relop (relational operator) is .EQ., ==, .NE., or /=, and variables e1 and e2 are type complexor real of any kind (precision). See Section "2.2 Compiler Options" for information on the -Ec option. The following is an example.

Example: Preferred decision

IF (ABS(A-0.1)-0.00001) 2,1,13 IF (ABS(B-C).LT.D) GO TO 4

In the constant expression defining a parameter, do not specify a constant expression whose evaluated result might be an denormalizednumber (see Section "5.1.3 Real and Complex Type Data"). Such a number may cause unexpected results.

6.3.3 Logical Data OperationsDo not assign a value other than the result of a logical expression (.TRUE. or .FALSE.) for logical data and treat that value as logical. Forexample, do not initialize logical variable with a hexadecimal number and treat the value as logical. Also, do not equivalence a logicaland integer variable and treat the integer as logical.

Example: Logical operation using other than logical values

LOGICAL A/Z'00000002'/,BINTEGER I/2/EQUIVALENCE (I,B)IF (A) GO TO 10IF (B) GO TO 20IF (A.AND.B) GO TO 30IF (A.OR.B) GO TO 40

Note:

In this case, do not expect the logical IF statement to execute the GO TO statement.

The result of a logical expression is TRUE or FALSE. Logical operators do not perform logical operations for each bit.

Example:

LOGICAL(1) L1,L2DATA L1/Z'FF'/L2=.NOT.L1

Note:

In this example, the result of L2 is not necessarily false.

6.3.4 Evaluation of OperationsThe result type of operation +, -, *, / , and ** between eight-byte integer and single-precision real or complex is different in Fortranversions.

For Fortran 95 and Fortran 2003 program, the result type of these operations is of type single-precision real or single precision-complex.

- 103 -

For FORTRAN66 and FORTRAN77 program, the result type of these operations is of type double precision-real or double-precisioncomplex.

6.3.5 Order of Evaluation of Data EntitiesThe order of evaluation of data entities connected by an operator is undefined.

For example, it is not allowed in a program to assume that one of F(X) and F(Y) should be evaluated previous to the other in the expressionF(X) + F(Y).

6.3.6 Evaluation of a part of ExpressionWhen an expression is evaluated, some parts of the expression may not be evaluated.

For example, in evaluating the logical expression L1.AND.L2 L2, L2 may not be evaluated if L1 is false because the value of the expressionis clearly false with out evaluating L2.

Therefore, such a function reference as a part of an expression must be written carefully.

6.3.7 Definition and Undefinition by Function ReferenceEvaluation of a function reference may not change the value of the data entity referred outside the function reference in the same statement.

For example, in the expression F(I).AND.I==1, evaluation of the function reference F(I) may not change the value of I.

Evaluation of the function reference in the logical expression of the logical IF statement can exceptionally change the value of the dataentity referred in the executable statement of the logical IF statement.

Additionally, evaluation of a function reference in a statement may not change the value of the common block element whose value mightaffect evaluation of any other function references in the same statement.

For example, it is not allowed evaluating the expression FA( ) + FB( ) where functions FA and FB are defined.

Example:

FUNCTION FA()COMMON /C/AFA=AENDFUNCTION FB()COMMON /C/AA=1FB=AEND

6.3.8 Variable enclosed by parenthesisAn actual argument is different in Fortran versions when a variable is enclosed by parenthesis in an actual argument of user's procedure.

In FORTRAN66 and FORTRAN77, the actual argument has the described variable.

In Fortran95 and Fortran 2003, the value of the variable is copied to the temporary area, and the actual argument has temporary area.

6.4 Assignment Statement

6.4.1 Assignment Statements with Overlapping Character PositionsIn FORTRAN66 and FORTRAN77, if the -Nf90move compiler option is not specified, overlapping data (in a storage area) may not bespecified on both sides of a character assignment statement. If the -Nf90move compiler option is specified, overlapping character data ispermitted.

In Fortran 95 and Fortran 2003, overlapping character data is permitted.

- 104 -

Example: In Fortran 95 and Fortran 2003, assignment of data with overlapping character positions

CHARACTER(LEN=5) CC='12345'C(1:4)=C(2:5)PRINT *,C ! output is '23455'

Note:

In FORTRAN66 and FORTRAN77, if the -Nf90move option is not specified, the assignment is not permitted.

6.4.2 Allocatable AssignmentWhen -Nalloc_assign compiler option is effective and left side of the assignment statement is an allocatable variable, diagnostic messagejwd2881i-i is output at compilation time and the allocatable assignment of the Fortran 2003 standard is operated at run time.

The execution performance decreases when the -Nalloc_assign compiler option is specified for the source program of the Fortran 95language specification.

The execution performance never decreases when the -Nnoalloc_assign compiler option is specified or the left side does not become anallocatable variable as the following example:

Example: No allocatable variable whose the section subscript Before: left part is allocatable variable

INTEGER,ALLOCATABLE::ARRAY(:)ALLOCATE(ARRAY(2))ARRAY=[1,2]END

After: left part is not allocatable variable

INTEGER,ALLOCATABLE::ARRAY(:)ALLOCATE(ARRAY(2))ARRAY(:)=[1,2]END

6.5 Control StatementsThis section describes considerations related to control statements.

6.5.1 GO TO StatementIf a statement label list is present and the destination of an assigned GO TO statement is a statement label other than one in the list, noerror is output during compilation. However, the program may not operate correctly.

To improve execution efficiency, specify all labels of statements that can be branch targets of the assigned GO TO statement.

6.5.2 Arithmetic IF StatementIf the arithmetic expression in an arithmetic IF statement is an integer expression and fixed-point overflow occurs, the next statementexecuted is unpredictable. For example, if the following IF statement is executed, the statement executed next could be 10, 20, or 30.

Example:

I=3 IF(2147483647+I)10,20,30

- 105 -

6.5.3 Logical IF StatementIf an expression in a logical IF statement has a value that does not conform to the internal logical data form, execution is unpredictable.For more information, see Section "5.1.2 Logical Type Data", and Section "6.3.3 Logical Data Operations".

Example:

LOGICAL TREQUIVALENCE(TR,JR)DATA JR/16/IF(TR) CALL S

Note:

The CALL statement may or may not be executed.

6.5.4 DO StatementIf there is a statement that branches into the range of a DO loop from outside the range of the DO loop, the system outputs a warningdiagnostic message. In this case, program operation is unpredictable.

For the extended range of a DO loop defined using FORTRAN66, the system outputs a diagnostic message as a warning, and the programoperates normally. However, the program should be corrected to ensure future compatibility.

6.5.4.1 DO Loop Iteration Count MethodThe following formula defines the DO loop iteration count r7 used in FORTRAN77, Fortran 95 and Fortran 2003.

r7 = MAX ( INT ( ( m2 - m1 + m3 ) / m3 ) , 0 )

In this formula, m1, m2, and m3 are the initial, terminal, and incrementation parameters of the DO loop. The body of the DO loop is notalways executed.

The following formula defines the DO loop iteration count r6 used in FORTRAN66.

r6 = MAX ( ( m2 - m1 ) / m3 + 1 , 1 )

For FORTRAN66, the body of the DO loop is executed at least once.

6.5.4.2 DO CONCURRENTIf a loop control of DO construct has CONCURRENT, the DO construct (DO CONCURRENT) is equal to DO construct withNORECURRENCE specifier in optimization control specifier. See Section "9.10.4 Optimization Control Specifiers" for information aboutNORECURRENCE specifier.

6.5.5 CASE StatementIf a case expression is an integer expression, the case value shall have the value that can be represent of case expression type and typeparameter.

Example:

INTEGER(2)::I...SELECT CASE (I)CASE (40000) ! THE CASE VALUE HAS THE VALUE BETWEEN -32768 TO ... ! 32767, BECAUSE THE CASE EXPRESSION IS OF TYPE ! TWO-BYTE INTEGEREND SELECT

- 106 -

6.5.6 PAUSE StatementThe PAUSE statement writes diagnostic message jwe0001a to the standard error file. If the message is sent to the terminal and the standardinput is from the terminal, execution of the program is suspended awaiting a response. Entering any standard input data at the terminalrestarts the program. In cases other than the above, the PAUSE statement is ignored and the program continues.

Example: PAUSE statement diagnostic messages

PAUSE jwe0001a pause

PAUSE n jwe0001a pause n

n: A decimal number of one to five digits

PAUSE c jwe0001a pause c

c: A character constant enclosed with apostrophes with a length not exceeding 255

6.5.7 STOP/ERROR STOP StatementThe STOP/ERROR STOP statement causes termination of the executing program, and writes the diagnostic message as follows to thestandard error output file.

Statement kind Stop-code specification Diagnostic message

STOPNone -

Yes jwe0002i

ERRO STOPNone

jwe0003iYes

-: No message

The stop-code of the STOP/ERROR STOP statement is reflected in the return code as follows if the environment variableFLIB_USE_STOPCODE has the value 1.

Statement kind Stop-code Return code

STOP

(Not specified) 0

scalar-int-initialization-exprThe least significant byte specified by a stop-code

(0 to 255)

scalar-default-char-initialization-expr 0

ERROR STOP

(Not specified) 1

scalar-int-initialization-exprThe least significant byte specified by a stop-code

(0 to 255)

scalar-default-char-initialization-expr 1

Example: STOP/ERROR STOP statement diagnostic messages

- STOP scalar-int-initialization-expr

STOP 123456 jwe0002i stop 123456

STOP -3 + 4 jwe0002i stop 1

- 107 -

- STOP scalar-default-char-initialization-expr

STOP 'NORMAL' jwe0002i stop NORMAL

STOP ('AB'//'CD')//'EF' jwe0002i stop ABCDEF

- ERROR STOP scalar-int-initialization-expr

ERROR STOP 123456 jwe0003i error stop 123456

ERROR STOP -3 + 2 jwe0003i error stop -1

- ERROR STOP scalar-default-char-initialization-expr

ERROR STOP 'ERROR' jwe0003i stop ERROR

ERROR STOP ('AB'//'CD')//'EF' jwe0003i stop ABCDEF

Example: Return code from the STOP/ERROR STOP statement Statement

kindStop-code Example sentence Diagnostic message Return

code

STOP

(Not specified) STOP - 0

scalar-int-initialization-expr

STOP -3 + 4 jwe0002i stop 1 1

STOP 400 jwe0002i stop 400 144

scalar-default-char-initialization-expr

STOP 'ABC' jwe0002i stop ABC 0

STOP ('AB'//'CD')//'EF' jwe0002i stop ABCDEF 0

ERRORSTOP

(Not specified) ERROR STOP jwe0003i error stop 1

scalar-int-initialization-expr

ERROR STOP 222jwe0003i error stop 222

222

ERROR STOP -3 + 2 jwe0003i error stop -1 255

scalar-default-char-initialization-expr

ERROR STOP 'ERROR'jwe0003i error stop ERROR

1

ERROR STOP ('AB'//'CD')//'EF'jwe0003i error stop ABCDEF

1

-: No message

6.5.8 ALLOCATE/DEALLOCATE StatementIf the STAT= specifier is specified for an ALLOCATE or DEALLOCATE statement, the following value is returned and executioncontinues.

Values except for 0 are runtime diagnostic message numbers. Note that you could not get a correct value if you use a 1-byte integer variableto save a return value.

- 108 -

Returnvalue

Meaning

0 Normal end

912 Available memory was insufficient when the ALLOCATE statement was executed.

1001 The allocatable variable specified for an ALLOCATE statement was already allocated.

1003 The allocatable variable specified for a DEALLOCATE statement was not allocated. The deallocation isignored.

1004 The pointer specified for a DEALLOCATE statement was not associated.

1005 The pointer specified for a DEALLOCATE statement is associated with a target that has not been createdby an ALLOCATE statement.

6.5.9 Simple Output List Enclosed in ParenthesesIn FORTRAN66, when a simple output list that is enclosed in parentheses and contains named constant is specified, it is output as a simpleoutput list enclosed in parentheses.

In FORTRAN77, Fortran 95 and Fortran 2003, it is output as a COMPLEX literal constant.

Example:

REAL A,BPARAMETER (A=1.0,B=2.0)WRITE (*,*) (A,B)END

In FORTRAN66, two REAL literals as a simple output list enclosed in parentheses are output as follows:

$ ./a.out 1.00000000 2.00000000$

In FORTRAN77, Fortran 95 and Fortran 2003, a COMPLEX literal constant is output as follows:

$ ./a.out (1.00000000,2.00000000)$

6.5.10 The Output Statement that starts by TYPE KeywordThe output statement that starts by TYPE keyword can be complied and executed in FORTRAN66 and FORTRAN77. In Fortran 95 andFortran 2003, the output statement that starts by TYPE keyword is interpreted as a TYPE statement of derived type definition or a TYPEtype declaration statement, and the diagnostic message at s level is output. Change the keyword TYPE to PRINT.

Example: The output statement that starts by TYPE keyword

A = 1.0TYPE*, A ! The diagnostic message at s level is output in Fortran 95 and Fortran 2003.END

6.6 ProceduresThis section describes various properties of procedures.

- 109 -

6.6.1 Statement Function ReferenceIf the types of actual and dummy arguments do not match in a statement function reference, the system outputs a diagnostic message. Theactual argument type is used for evaluation. If an arithmetic or logical expression is used for the actual argument, however, the actualargument type must follow the expression conversion rules.

In the following example, the actual arguments are double-precision real for statement function SF. However, statement function ESFcauses an error.

Example: Statement function reference

REAL(8) X,Y,ZLOGICAL L,M,N,ESFSF(I,J,K)=I**2+J**3+K**4ESF(L,M,N)=.NOT. L .AND. M .AND. N:XX=3.0E1+SF(X,Y,Z) ! Actual arguments of SF are: ! double-precision realL=X==Y .AND. ESF(X,Y,Z) ! These arguments not allowed ! by conversion rules

6.6.2 An Intrinsic Function that Specifies EXTERNAL AttributeIn FORTRAN77, Fortran 95 and Fortran 2003, if the EXTERNAL attribute is specified to an intrinsic function name, the intrinsic functioncharacteristic is lost, and the name is processed as a user-defined external procedure or block data program unit.

In FORTRAN66, if a basic external function name is specified in an EXTERNAL statement, the name is processed as a basic externalfunction.

Note:

The FORTRAN77, Fortran 95 and Fortran 2003 intrinsic function is the generic name for a FORTRAN66 basic external or intrinsicfunctions.

6.6.3 REAL and CMPLX Used as Generic NamesThe generic function REAL converts the data to type real in FORTRAN77, Fortran 95 and Fortran 2003, but it only fetches the real partfrom a complex argument in FORTRAN66.

In Fortran 95 and Fortran 2003, if the first argument is of type complex and the second argument is not present in the generic functionREAL, the result type parameter is the kind type parameter of the first argument.

The generic function CMPLX converts the date to type complex in FORTRAN77, Fortran 95 and Fortran 2003, but it only creates acomplex number in FORTRAN66.

6.6.4 Intrinsic ProceduresIn FORTRAN77, Fortran 95 and Fortran 2003, the following names are interpreted as intrinsic functions. In FORTRAN66, the samenames are treated as user-defined functions and not as intrinsic functions (basic external functions).

ICHAR, CHAR, ANINT, DNINT, QNINT, NINT, IDNINT, IQNINT, DPROD, QPROD, LEN, INDEX, DASIN, QASIN, DACOS, QACOS, LGE, LGT, LLE, LLT, NOT, IAND, IOR, IEOR, ISHFT, IBSET, IBCLR, BTEST

In Fortran 95 and Fortran 2003, the following names are interpreted as intrinsic procedures. In FORTRAN66 and FORTRAN77, the samenames are treated as user-defined procedures.

ACHAR, ADJUSTL, ADJUSTR, ALL, ALLOCATED, ANY, ASSOCIATED, BIT_SIZE, CEILING, COMMAND_ARGUMENT_COUNT , COUNT, CPU_TIME, CSHIFT, DATE_AND_TIME, DIGITS, DOT_PRODUCT, EOSHIFT, EPSILON, EXPONENT, FLOOR, FRACTION, HUGE, GET_COMMAND, GET_COMMAND_ARGUMENT, GET_ENVIRONMENT_VARIABLE, IACHAR, IBITS, ISHFTC, KIND, LBOUND, LEN_TRIM, LOGICAL, MATMUL, MAXEXPONENT, MAXLOC, MAXVAL, MERGE, MINEXPONENT, MINLOC, MINVAL, MODULO, MOVE_ALLOC, MVBITS, NEAREST, NEW_LINE , NULL , PACK, PRECISION, PRESENT, PRODUCT, RADIX, RANDOM_NUMBER, RANDOM_SEED, RANGE, REPEAT, RESHAPE, RRSPACING, SAME_TYPE_AS, SCALE, SCAN, SELECTED_CHAR_KIND,

- 110 -

SELECTED_INT_KIND, SELECTED_REAL_KIND, SET_EXPONENT, SHAPE, SIZE, SPACING, SPREAD, SUM, TINY, SYSTEM_CLOCK, TRANSFER, TRANSPOSE, TRIM, UBOUND, UNPACK, VERIFY

If the -Nobsfun compiler option is specified, the following names are interpreted as intrinsic functions.

AIMAX0, AJMAX0, I2MAX0, IMAX0, JMAX0, IMAX1, JMAX1, AIMIN0, AJMIN0, I2MIN0, IMIN0, JMIN0, IMIN1, JMIN1, FLOATI, FLOATJ, DFLOTI, DFLOTJ, IIABS, JIABS, I2ABS, IIDIM, JIDIM, I2DIM, IIFIX, JIFIX, IFIX, INT1, INT2, INT4, INT8, IINT, JINT, ININT, JNINT, IIDNNT, I2NINT, JIDNNT, IIDINT, JIDINT, IMOD, JMOD, I2MOD, IISIGN, JISIGN, I2SIGN, BITEST, BJTEST, IIBCLR, JIBCLR, IIBITS, JIBITS, IIBSET, JIBSET, IBCHNG, ISHA, ISHC, ISHL, IIAND, JIAND, IIEOR, JIEOR, IIOR, JIOR, INOT, JNOT, IISHFT, JISHFT, IISHFTC, JISHFTC, IZEXT, JZEXT, IZEXT2, JZEXT2, JZEXT4, VAL

6.6.5 Intrinsic Function Names with the Type DeclaredIn FORTRAN77, Fortran 95 and Fortran 2003, even if an intrinsic function is declared with a different type, the intrinsic functioncharacteristic is not lost.

In FORTRAN66, if an intrinsic function is declared with a different type, the intrinsic function characteristic is lost. In addition, the intrinsicfunction is treated as a user-defined function whenever it is used.

6.6.6 Service RoutinesIf compiler option -AU is specified, a service subroutine name must be entered in lowercase characters.

The validity of the number and type of arguments is checked if the module SERVICE_ROUTINES that includes all service routinesexplicit interface is used. There are following notes to use the SERVICE_ROUTINES:

- Specify ONLY option and service routines that only using in program because SERVICE_ROUTINES includes all service subroutinesand service functions explicit interface. For example,

USE SERVICE_ROUTINES,ONLY:ACCESS

- If compiler option -AU is specified, module name SERVICE_ROUTINES and each service routine name in USE statement must bein lowercase characters.

- If compiler option -CcdII8, -CcI4I8, -CcdRR8, -CcR4R8, -CcdLL8, -CcL4L8, -Ccd4d8, -Cca4a8, -Ad or -Aq is specified, thecorresponding types of argument of service routines is changed. Specify the correspondence runtime option -Lb, -Li, or -Lr.

The validity of the number and type of arguments is not checked if the module SERVICE_ROUTINES is not used. For about module anduse a module, see "Chapter 10 Fortran Modules".

If the -Nmallocfree compiler option is specified, the MALLOC and the FREE service routines are interpreted as intrinsic procedures.

If the compiler option -NRtrap and the runtime option -Wl,-i or the environment variable FLIB_EXCEPT=u are specified, DVCHK servicesubroutine and OVERFL service subroutine return the value which shows that an exception doesn't occur.

6.6.7 The returned values of the intrinsic functionWhen negative zero is specified for the argument, the following cases become the different returned value between Fortran2003 andFortran95 or before.

- ATAN2(Y,X) : When X < 0 and Y is negative zero, the returned value is -PI in Fortran2003 and PI in Fortran95 or before.

- ATAN2D(Y,X) : When X < 0 and Y is negative zero, the returned value is -180.0 in Fortran2003 and 180.0 in Fortran95 or before.

- ATAN2Q(Y,X) : When X < 0 and Y is negative zero, the returned value is -2.0 in Fortran2003 and 2.0 in Fortran95 or before.

- LOG(X) : When X is complex with REAL(X) < 0 and the imaginary part of X is negative zero, the imaginary part of the returnedvalue is -PI in Fortran2003 and PI in Fortran95 or before.

- SQRT(X) : When X is complex with REAL(X) < 0 and the imaginary part of X is negative zero, the imaginary part of the returnedvalue is negative value in Fortran2003 and positive value in Fortran95 or before.

- 111 -

Chapter 7 Input/output ProcessingThis section describes the basic properties of Fortran records and files and the Fortran input/output statements used to process files. Featuresthat are ISO standard Fortran are not discussed in general, except to describe their specific characteristics in the Fujitsu Fortran system.Unless indicated otherwise, "file" means external file in this section.

7.1 FilesInput/output statements process internal and external files. Internal files are stored in main memory.

Internal files consist of scalar character variables, character array elements, character arrays, or substrings.

Internal input/output statements may read or write internal files. External files are stored on external devices and are connected forsequential, direct, and stream access.

Fortran programs process the following types of files:

- Standard input files (stdin)

- Standard output files (stdout)

- Standard error output files (stderr)

- Regular files

Files other than standard files are called regular files. Either sequential, direct, or stream access methods can be used to access regularfiles. Only the sequential access method can be used to access standard files. Unformatted input/output statements may not be used toaccess standard files.

7.1.1 Old and New FilesNew files may be created when an executable program is started or at any time during its execution.

Old and new files may be accessed when connected to a unit number specified in an input/output statement.

7.1.2 File ConnectionA file and unit must be connected to execute a data transfer or file positioning input/output statement.

7.1.2.1 ConnectionFiles are connected in two ways:

- Dynamic connection (during execution of a program)

- Preconnection

Dynamic connection is performed when an OPEN, READ, WRITE, BACKSPACE, ENDFILE, or PRINT statement is executed.

However, dynamic connection for direct or stream access may be performed only by an OPEN statement.

Preconnection is performed by an external specification for a Fortran program before an executable program is started.

The standard input file (stdin, unit 5), the standard output file (stdout, unit 6), and the standard error output files (stderr, unit 0) are alwayspreconnected. Executing an OPEN statement with a FILE= specifier disconnects a standard file.

7.1.2.2 File disconnectionFiles are disconnected in the following cases.

- Execution of an CLOSE statement

- Termination of the execution program

- Execution of an OPEN statement with different file name to the same unit number.

- 112 -

The following processes are done in the disconnection of the file.

- Termination of the connection of specified unit to an external file

- Deletion or keeping of the file that is connected to the specified unit

A file that has been disconnected may be connected by an OPEN statement.

The only means of connecting a file that has been disconnected is by appearance of its name in an OPEN statement.

There may be no means if reconnecting an unnamed file once it is disconnected.

7.2 Unit Numbers and File ConnectionUnits and files are connected as follows:

- Execute an OPEN statement with a filename specified.

- Use a environment variable outside the program to pre-connect a unit and file.

7.2.1 Priority of Unit and File ConnectionThe way of connecting it in the case of standard input, standard output and the standard error output file and that order of priority areshown in the following.

How to specify Explanation Priority

An OPEN statement with FILE=specifier Connection by an OPEN statement with FILE=specifier High

Runtime options Connection by the runtime options.

-pu_no (standard output)

-ru_no (standard input)

-mu_no (standard error output)

Default value of Fortran system Connection by Fortran System.

Standard output (unit number 6)

Low

- 113 -


Standard input (unit number 5)

Standard error output (unit number 0)

The way of connecting it and that order of priority are shown in the following in the case as the named file and the unnamed file.


An OPEN statement with FILE=specifier Connection by an OPEN statement with FILE=specifier High

The environment variable Connection by the environment variable (fuxx). Low

7.2.2 Environment variables to Connect Units and FilesEnvironment variables may be used to connect units and files. How to specify file name by using an environment variable is shown in thefollowing.

Name of Environment variable Value

fuxx File-name

fu: Fixed

xx: Unit number (00 to 2147483647)

File-name: Filename (full) pathname or filename

Unit numbers 0, 5, and 6 (set by runtime options -m, -r, and -p) are preconnected for error output, standard input, and standard output,respectively. These unit numbers may not be connected using a environment variable.

Environment variable file connections are disabled when a file is closed and remain disabled thereafter. The input/output statements thatmay close any of the standard files are as follows:

- CLOSE statement

- OPEN statement with a file specified

Example 1: Connect "test.data" to unit number 1

Environment variable setting:

$export fu01="test.data"

Example 2: Disabled connect by environment variable

Fortran program:

OPEN(1,FILE='b.file') ...

Commands:

$ export fu01=a.file$ ./a.out

When a.out is executed, file connection by environment variable (connection to a.file) is disabled, and the filename specified an OPENstatement (connection to b.file) is enabled.

7.3 RecordsFortran records contain data in internal machine format or character strings. One Fortran record does not always correspond to one physicalrecord. The Fortran record types are as follows:

- Formatted (see Section 7.3.1 Formatted Fortran Records)

- 114 -

- Unformatted (see Section 7.3.2 Unformatted Fortran Records)

- List-Directed (see Section 7.3.3 List-Directed Fortran Records)

- Namelist (see Section 7.3.4 Namelist Fortran Records)

- Endfile (see Section 7.3.5 Endfile Records)

- Binary (see Section 7.3.6 Binary Fortran Records)

- Stream (see Section 7.3.7 STREAM Fortran Records)

Fortran record characteristics are shown below.

Fortran record

typeInput/output statement Record format One Fortran record

Formatted

Formatted sequential Undefined lengthOne logical record

Formatted direct

Fixed lengthInternal file

One scalar character variable, arrayelement, or substring.

For an array, each element is aFortran record.

UnformattedUnformatted sequential Variable length

One logical recordUnformatted direct Fixed length

List-directed

List-directed, Print statement Undefined length One logical record

Internal file Fixed length



Namelist

Namelist Undefined lengthOne logical record from the&namelist name up to / or &end.

Internal file Fixed length



EndfileInput/output statement otherthan direct

----The last Fortran record does nothave the length attribute.

BinaryUnformatted sequential Fixed length(*1)

One logical recordUnformatted direct Fixed length(*1)

Streamformatted stream Undefined length One logical record

Unformatted stream Fixed length (*1) One logical record

*1. The Fortran record length is determined according to the size of data for which input/output is to be executed.

Fortran records exist as logical records on external devices. An undefined length record consists of a Fortran record and the line-feedcharacter (\n). A fixed length record consists of a Fortran record. A variable length record consists of a Fortran record and 8 bytes (a4-byte length before and after the Fortran record).

7.3.1 Formatted Fortran RecordsFormatted Fortran records can consist of any character string. Formatted Fortran records are accessed by formatted sequential, formatteddirect, and internal file input/output statements.

- 115 -

7.3.1.1 Formatted Sequential RecordFiles controlled by formatted sequential input/output statements have an undefined length record format.

One Fortran record corresponds to one logical record. The length of the undefined length record depends on the Fortran record to beprocessed. The max length may be assigned in OPEN statement RECL= specifier.

The line-feed character (\n) terminates the logical record. If the $ edit descriptor or \ edit descriptor is specified for the format of the formatof the formatted sequential output statement, the Fortran record does not include the line feed character (\n).

The relationship between formatted Fortran records and the undefined length record format is shown below.

Fortran program:

INTEGER,DIMENSION(20) :: IA INTEGER,DIMENSION(10) :: IB INTEGER,DIMENSION(30) :: IC WRITE(2,10) IA,IB 10 FORMAT(20I4,/,10I4) WRITE(2,20) IC 20 FORMAT(30I4)

Fortran records:

7.3.1.2 Formatted Direct RecordFiles processed by formatted direct input/output statements have a fixed length record format. One Fortran record corresponds to onelogical record. The length of the logical record must be assigned in the OPEN statement RECL= specifier. If the Fortran record is shorterthan the logical record, the remaining part is padded with blanks. The length of the Fortran record must not exceed the logical record. Thisfixed length record format is unique to Fortran.

The relationship between formatted Fortran records and the fixed length record format is shown below.

Fortran program:

INTEGER,DIMENSION(20) :: IA INTEGER,DIMENSION(10) :: IB OPEN(UNIT=3,ACCESS='DIRECT',RECL=80,FORM='FORMATTED') WRITE(3,FMT=10,REC=1) IA,IB 10 FORMAT(20I4)

- 116 -

Fortran record:

7.3.1.3 Internal File RecordThe fixed length record are processed by internal file input/output statements. The records are stored in internal files. Internal files thatare a scalar character variable, character array element or substring contain only one Fortran record. The length of the Fortran record mustnot exceed the size of the scalar character variable, character array element, or substring. When an internal file is a character array, thelength of a Fortran record must not exceed the size of an array element of the character array. If the length of the Fortran record is shorterthan the size of the scalar character variable, character array element or substring, the remaining part is padded with blanks. For input, thelengths must be equal.

The relationship between formatted Fortran records and internal files is shown below.

Fortran program:

CHARACTER(LEN=20) C(3)

INTEGER,DIMENSION(5) :: IA

INTEGER,DIMENSION(3) :: IB

INTEGER,DIMENSION(2) :: IC

WRITE(C,10) IA,IB,IC 10 FORMAT(5I4)

Fortran record (Internal file: C):

7.3.2 Unformatted Fortran RecordsAn unformatted Fortran record consists of a string of values (character or noncharacter data, and a string can be empty). The length dependson the input/output source program statement (zero length can be used).

Both unformatted sequential and unformatted direct input/output statements can use unformatted Fortran records.

7.3.2.1 Unformatted Sequential RecordFiles processed using unformatted sequential input/output statements have a variable length record format.

One Fortran record corresponds to one logical record. The length of the variable length record depends on the length of the Fortran record.The length of the Fortran record includes 4 bytes added to the beginning and end of the logical record. If the runtime option -Wl,-Lu isspecified, the size of the field to put the length has 8 bytes. The max length may be assigned in the OPEN statement RECL= specifier.The beginning area is used when an unformatted sequential statement is executed. The end area is used when a BACKSPACE statement

- 117 -

is executed. This variable length record format is unique to Fortran. The relationship between unformatted Fortran records and the variablelength record format is shown below.

Fortran program:

INTEGER,DIMENSION(10) :: IA,ICINTEGER,DIMENSION(20) :: IBWRITE(2) IA,IBWRITE(2) IC

Fortran record (the runtime option -Wl,-Lu is not specified):

L: Length of Fortran record

Fortran record (the compiler option is valid and the runtime option -Wl,-Lu is specified):

L: Length of Fortran record

The Fortran record that length is 2G bytes (2147483648 bytes) or more is input/output by dividing the record. The length of each dividedFortran record is from 1 to 2147483647 bytes.

- 118 -

The value set to the top and the end of a logical record is as follows because the divided Fortran records are considered single Fortranrecord.

Divided

Fortran recordThe value of top of logical record The value of end of logical record

First

A negative value that the absolute value is equalto the length of the divided Fortran record is set.

(-2147483647 is set)

The length of the divided Fortran record is set.

(2147483647 is set)

Middle

A negative value that the absolute value is equalto the length of the divided Fortran record is set.

(-2147483647 is set)

A negative value that the absolute value is equal tothe length of the divided Fortran record is set.

(-2147483647 is set)

LastThe length of the divided Fortran record is set. A negative value that the absolute value is equal to

the length of the divided Fortran record is set.

The relationship between unformatted Fortran records that length is 2G bytes or more and the variable length record format is shownbelow.

Fortran program:

CHARACTER(LEN=2000000000) :: CH1,CH2,CH3OPEN(10,FILE="work.data",FORM="UNFORMATTED")WRITE(10) CH1,CH2,CH3

Fortran record:

L: Length of Fortran recordRF: Divided Fortran record (First)RM: Divided Fortran record (Middle)RL: Divided Fortran record (Last)

- 119 -

7.3.2.2 Unformatted Direct RecordFiles processed by unformatted direct input/output statements have a fixed length record format. One Fortran record can correspond tomore than one logical record. See Section 7.3.1.2 Formatted Direct Record for details.

The record length must be assigned in the OPEN statement RECL= specifier. One Fortran record can consist of more than one logicalrecord. If the Fortran record terminates within a logical record, the remaining part is padded with binary zeros. If the length of the Fortranrecord exceeds the logical record, the remaining data goes into the next record. This fixed length record format is unique to Fortran.

7.3.3 List-Directed Fortran RecordsList-directed input/output statements, print statement, or internal file input/output statements are used to process list-directed Fortranrecords. List-directed Fortran records consist of data items and value separators. The data items are input/output character strings. Thelength of the Fortran record depends on the number and type of items. Data items from the beginning of a logical record up to the end ofthe input are handled as one Fortran record. Files processed using list-directed input/output statements have an undefined length recordformat or a fixed length record format. See Section 7.3.1.1 Formatted Sequential Record for details. See Section 7.3.1.3 Internal FileRecord for details.

7.3.4 Namelist Fortran RecordsNamelist input/output statements and internal input/output statements access namelist Fortran records.

Namelist Fortran records consist of data items (character strings specified by a namelist name) from the &namelist name up to / or&END. The correspondence between namelist Fortran records and logical records is the same as for Fortran records handled using list-directed input/output statements. Files accessed by namelist input/output statements have an undefined length record format. See Section7.3.1.1 Formatted Sequential Record for details. See Section 7.3.1.3 Internal File Record for details.

7.3.5 Endfile RecordsAn endfile record must be the last record of a file connected for sequential access. Endfile records do not have a length attribute.

The ENDFILE statement writes endfile records in files connected for sequential access. After at least one WRITE statements is executed,endfile records are output under the following conditions:

- A REWIND statement is executed.

- A BACKSPACE statement is executed.

- A CLOSE statement is executed.

7.3.6 Binary Fortran RecordsA binary Fortran record consists of a string of values (character or noncharacter data). The length of the binary Fortran record depends onthe input/output list (it may be zero length). One Fortran record corresponds to one logical record. An unformatted sequential access oran unformatted direct access input/output statement may access binary Fortran records. See Section 7.3.1.2 Formatted Direct Record fordetails.

7.3.7 STREAM Fortran RecordsA STREAM Fortran record is composed of the file storage units, and its storage position can be uniquely identified using positive integers.

STREAM Fortran records can be accessed by stream input/output statements.

7.3.7.1 Formatted Stream RecordFiles controlled by formatted stream input/output statements have an undefined length record format.

One Fortran record corresponds to one logical record. The length of the undefined length record depends on the Fortran record to beprocessed. The line-feed character (\n) terminates the logical record. If the $ edit descriptor or \ edit descriptor is specified for the formatof the formatted stream output statement, the Fortran record does not include the line feed character (\n).

- 120 -

Fortran program:

OPEN(10, FILE='x', ACCESS='stream', FORM='formatted')DO I=1,10 INQUIRE(10,POS=N) WRITE(10,'(I4)',POS=N) IENDDOCLOSE(10)

Fortran record:

7.3.7.2 Unformatted Stream RecordFiles controlled by unformatted stream input/output statements have a fixed length record format.

The length of the undefined length record depends on the Fortran record to be processed.

The relationship between unformatted Fortran records and the fixed length record format is shown below.

Fortran program:

OPEN(10, FILE='x', ACCESS='stream', FORM='unformatted')DO I=1,10WRITE(10,POS=4*I ) IENDDOCLOSE(10)

Fortran record:

7.4 OPEN Statement

7.4.1 OPEN Statement SpecifiersThese topics explain the parameters that can be specified in an OPEN statement. The parameters are:

- FILE= Specifier

- STATUS= Specifier

- RECL= Specifier

- ACTION= Specifier

- BLOCKSIZE= Specifier

- CONVERT= Specifier

- CARRIAGECONTROL= Specifier

- 121 -

The character strings that can be specified with STATUS= , ACCESS= , FORM= , BLANK= , ACTION= , PAD= , POSITION= , DELIM= ,ASYNCHRONOUS= , DECIMAL= , ROUND= , SIGN= , CARRIAGECONTROL= , and CONVERT= specifiers depend on theparameter.

The characters in the specifier may be any mixture of upper and lower case; however, ignoring case, they must match a legal specifierexactly. Blanks at the beginning are ignored. If the result is not an allowed value, diagnostic message jwe0086i-e is output, and the parameteris ignored. A specifier may contain trailing blanks.

Example values for the OPEN statement ACCESS= specifier are shown below. The other parameters are handled similarly.

Example:

Correct character expressions are:

OPEN(10,ACCESS='SEQUENTIAL')OPEN(10,ACCESS='sequential')OPEN(10,ACCESS='SeQuEnTiAl')OPEN(10,ACCESS='SEQUENTIAL ')OPEN(10,ACCESS=' SEQUENTIAL')

Incorrect character expressions are:

OPEN(10,ACCESS='SEQUENTIALAB')OPEN(10,ACCESS='SEQUEN')OPEN(10,ACCESS='S EQUENTIAL')OPEN(10,ACCESS=' ')

7.4.1.1 FILE= SpecifierThe FILE= specifier specifies a file. Blanks at the beginning of the FILE= specifier are ignored. If a blank is detected in the characterstring, the value of the string up to the blank is read.

Filename specified in a FILE= specifier is shown below, where it is assumed that the current directory is /home/v1954.

Interpreted as /home/v1954/file.data1 are:

OPEN(10,FILE='/home/v1954/file.data1')OPEN(10,FILE='file.data1')OPEN(10,FILE=' file.data1')

Interpreted as /home/c1121/file.data1 is:

OPEN(10,FILE='../c1121/file.data1')

Filename is specified with the FILE= specifier and the STATUS= specifier. The tables below list the relationship between FILE= andSTATUS= specifiers.

FILE= specifierSTATUS= specifier

NEW, OLD, SHR. REPLACE SCRATCH UNKNOWN

FILE= specifierThe system treats the specifiedvalue as a named file.

An Error.The system treats the specifiedvalue as NEW or OLD,depending on the file status.

No FILE= specifier(no environmentvariable)

The system treats thespecification as a named filewith file name fort.nn (nn: unitnumber).

The system treats thespecification as an unnamedfile.

The system treats thespecification as a named filewith file name fort.nn (nn: unitnumber).

No FILE= specifier(environmentvariable allocated)

The system treats the file namespecified by the shell variableas a named file.

An Error.The system treats the file namespecified by the environmentvariable as a named file.

Notes:

- 122 -

- If a FILE= specifier is not specified, processing depends on whether the units and file are connected by an environment variable (SeeSection 7.2 Unit Numbers and File Connection).

- If a name is not given, the name used is 'tFnnn', where nnn is a unique number. See "tempnam(3s)".

7.4.1.2 STATUS= SpecifierThe STATUS= specifier must correspond to the file status. That is, NEW or SCRATCH must not be specified for an existing file, and ifOLD or SHR is specified, the file must exist. When UNKNOWN is specified, the Fortran system determines the file status. If the STATUS=specifier is omitted, UNKNOWN is used by default.

If UNKNOWN and the FILE= specifier are specified, and the file does not already exist, the status is defined as NEW. If UNKNOWNand the FILE= specifier are specified, and the file does exist, the status is defined as OLD. If UNKNOWN and the environment variableare specified, and a FILE= specifier is not specified, the status is defined as OLD. If UNKNOWN is specified, and a FILE=specifier andthe environment variable are not specified, the status is defined as SCRATCH. The FILE= specifier may be present when STATUS isSCRATCH.

The tables below list the relationship between file and STATUS= specifier.

SHR is the same as OLD. If SHR is specified, the file must exist, and a FILE= specifier must be specified at the same time. When SHRis specified, the connected file can be shared with other executable programs.

Relationship between File and STATUS= specifier

STATUS= specifier File

NEW Specified for a file that does not exist.

OLD, SHR Specified for a file that exists.

SCRATCH Specified for a file that does not exist. The file is deleted when the program isfinished or the file is closed.

UNKNOWN Specified if the file status is unknown. The Fortran system determines the statusas follows:

- If the FILE= specifier is specified, the status is NEW if the file does notexist. The status is OLD if the file exists.

- If no FILE= specifier is specified, the status is SCRATCH if the file wasnot assigned using an environment variable. The status is OLD if the filewas assigned using an environment variable.

REPLACE If the file does exist, the file is deleted, and a new file is created with the samename.

7.4.1.3 RECL= SpecifierThe RECL= specifier specifies the length of each Fortran record in files in which DIRECT is specified by the ACCESS= specifier. AnOPEN statement with RECL= specified must be executed before any direct input/output statements are executed for that file. The valuespecified for an existing file must be the same as for each record in the file.

The RECL= specifier indicates the maximum length of a Fortran record in files with SEQUENTIAL specified by the ACCESS= specifier.If the RECL= specifier is omitted, 2147483647 is used by default.

The value of the RECL= specifier must be positive.

The RECL= specifier must not be specified for a file connected for stream access.

The tables below list the relationship between ACCESS= specifier and RECL= specifier.

Relationship between ACCESS= specifier and RECL= specifier

- 123 -

ACCESS= specifier with RECL= specifier without RECL= specifier

in Fortran95 and Fortran2003

in FORTRAN77 andFORTRAN66

SEQUENTIAL Sequential access Sequential access (*1) Sequential access

DIRECT Direct access Direct access Error

STREAM Error Error Stream access

None Sequential access Direct access (*2) Sequential access

*1: The diagnostic message (jwe0097i-w) is generated, and RECL= specifier is ignored.

*2: The compiler message (jwd1449i-i) is generated.

7.4.1.4 ACTION= SpecifierWhen READ is specified, "r" is the permission of the file; that is, WRITE and PRINT statements may not be executed for the file.

When WRITE is specified, "w" is the permission of the file; that is, READ, BACKSPACE, and REWIND statements cannot be executedfor the file. For direct access, READWRITE or BOTH is assumed even if WRITE is specified.

BOTH is the same as READWRITE.

When READWRITE or BOTH is specified, "r+" or "w+" is the permission of the file; that is, all input/output is allowed for the file. Whena file with read only permission is opened, READ must be specified. Therefore, when the WRITE statement is executed, a diagnosticmessage is output even if READWRITE or BOTH is specified. If the ACTION= specifier is omitted, READWRITE is used by default.

The tables below list the relationship between ACTION= specifier and file status.

Relationship between ACTION= specifier and Files

Access method File STATUS ACTION=specifier Open mode Notes

Sequential

NEW

READ Error

WRITE w

READWRITE orBOTH

w+

OLD, SHR

READ r

WRITE w

READWRITE orBOTH

r+ (*1)

SCRATCH

READ r ACTION='READ' isnot allowed fortemporary files andnew files.

WRITE w

READWRITE orBOTH

w+ (*1)

REPLACE

READ r

WRITE w+

READWRITE orBOTH

W+ (*2)

DirectNEW

READ Error

WRITE w

READWRITE orBOTH

w+

OLD, SHR READ r

- 124 -

Access method File STATUS ACTION=specifier Open mode Notes

WRITE r+

READWRITE orBOTH

r+ (*2)

SCRATCH

READ r ACTION='READ' isnot allowed fortemporary files andnew files.

WRITE w+

READWRITE orBOTH

r+ (*2)

REPLACE

READ r

WRITE w+

READWRITE orBOTH

w+ (*2)

Stream

NEW

READ Error

WRITE w

READWRITE orBOTH

w+

OLD, SHR

READ r

WRITE w

READWRITE orBOTH

r+ (*1)

SCRATCH

READ rACTION='READ' isnot allowed fortemporary files andnew files.

WRITE w

READWRITE orBOTH

w+ (*1)

REPLACE

READ r

WRITE w

READWRITE orBOTH

w+ (*2)

*1: A file with the read-only attribute is opened in 'r' mode.

*2: A file with the read-only attribute causes an error.

7.4.1.5 BLOCKSIZE= SpecifierThe BLOCKSIZE= specifier specifies the sequential input/output buffer size in bytes. The default buffer size is 8M bytes.

7.4.1.6 CONVERT= SpecifierWhen LITTLE ENDIAN or IBM is specified in the CONVERT= specifier, no asynchronous data transfer is performed when anasynchronous input/output statement is executed on that unit.

7.4.1.7 CARRIAGECONTROL= SpecifierWhen STREAM has been specified in the ACCESS= specifier, FORTRAN cannot be specified in the CARRIAGECONTROL= specifier.

7.4.2 Executing an OPEN Statement on a Connected FileIf the file connected to a unit is the same as the file to be connected, only BLANK=, DELIM=, PAD=, DECIMAL=, ROUND=, SIGN=,ERR=, and IOSTAT= specifiers may have values different from those currently in effect. If the other specifiers have different value, the

- 125 -

specified value is ignored in FORTRAN66 and FORTRAN77. In Fortran 95 and Fortran 2003, the runtime message (jwe1115i-w) is outputand the specified value is ignored.

In addition, if the value of the STATUS= specifier is other than OLD, in the Fortran 2003 language specifications the diagnostic messagejwe1115i-w is output, and the status of the external device that was connected immediately beforehand takes effect.

Example:

$ cat ap.f OPEN(10,FILE='XX.DAT',POSITION='APPEND') WRITE(10,*)123 OPEN(10,FILE='XX.DAT',POSITION='REWIND') WRITE(10,*)456 END

In FORTRAN66 and FORTRAN77:

$ ./a.out$ more XX.DAT 123 456$

In Fortran95 and Fortran 2003:

$ ./a.outjwe1115i-w line 7 The value of POSITION specifier in an OPEN statement is different from that currently in effect (unit=10). error occurs at MAIN__ line 7 loc 0000000000400b51 offset 0000000000000041 MAIN__ at loc 0000000000400b10 called from o.s.taken to (standard) corrective action, execution continuing.error summary (Fortran)error number error level error count jwe1115i w 1total error count = 1$ more XX.DAT 123 456$

7.4.3 Input/output of Fortran RecordsIf an input/output statement is executed that is inconsistent with the file properties of the connection (open), the specifications in the input/output statement are used in FORTRAN66 and FORTRAN77. In Fortran 95 and Fortran 2003, the runtime message (jwe0113i-e) is outputand the input/output statement is ignored.

Example:

$ cat fm.f OPEN(10,FILE='XX.DAT',FORM='UNFORMATTED') WRITE(10,*)123 END


$ ./a.out$ more XX.DAT 123$

In Fortran 95 and Fortran 2003:

$ ./a.outjwe0113i-e line 2 A(an) formatted I/O statement cannot be executed for a unit connected to unformatted (unit=10).

- 126 -

error occurs at MAIN__ line 2 loc 0000000000400b42 offset 0000000000000032 MAIN__ at loc 0000000000400b10 called from o.s.taken to (standard) corrective action, execution continuing.error summary (Fortran)error number error level error count jwe0113i e 1total error count = 1$ more XX.DAT$

7.4.4 Default Value of ACCESS= Specifier with RECL=SpecifierIn FORTRAN66 and FORTRAN77, if RECL= is specified and ACCESS= is not specified in an OPEN statement, ACCESS='DIRECT'is used by default. In Fortran 95 and Fortran 2003, if RECL= is specified and ACCESS= is not, ACCESS='SEQUENTIAL' is used bydefault.

Example: OPEN statement with RECL= and with no ACCESS= specifier:

OPEN (1, RECL=80)

7.5 CLOSE Statement STATUS= SpecifierThis section explains the STATUS= specifier used in the CLOSE statement. See Section 7.7.1 Control Information for Input/output DataTransfer Statements for information on the other specifiers.

7.5.1 STATUS= SpecifierThe STATUS= specifier indicates whether to keep or delete the file connected to the unit number when it is closed. This specifier and theOPEN statement STATUS= specifier determine whether the file is retained or deleted. The table below lists the relationship betweenSTATUS= specifiers of OPEN and CLOSE statements.

Relationship between STATUS= specifiers of OPEN and CLOSE statements

OPEN statement CLOSE statement

KEEP/FSYNC DELETE Default specification

NEW Keep the file after

CLOSE statement

execution.

Delete the file after

CLOSE statement

execution.

Keep the file after

CLOSE statement

execution.

OLD, SHR,

REPLACE

SCRATCH Delete an unnamed

file after CLOSE statementexecution.

Delete the file after

CLOSE statement

execution.

Note:

If UNKNOWN is specified in the OPEN statement, the file status is based on whether the file is handled as NEW, OLD, or SCRATCH.

In case the 'FSYNC' is specified to the CLOSE statement, the data which is in memory for the file is synchronized with storage device byinvoking the fsync(2) system call only if the following conditions are met:

1. The connected file is a named regular file.

2. The file listed in 1. above exists.

3. The file listed in 1. above is being used.

4. File system supports synchronous operation by invoking the fsync(2) system call.

If the condition listed in 1. above is not met, an error occurs when FSYNC is specified to CLOSE statement.

If the condition listed in 2. or 3. above is not met, the fsync(2) system call is not invoked, and processing is continued.

- 127 -

If the condition listed in 4. is not met, an error occurs.

7.6 INQUIRE Statement

7.6.1 INQUIRE Statement ParametersThe following table indicates the way the system treats returned INQUIRE values, depending on whether the length of the returned charactervalue is less than the scalar character variable name or array element name to be set.

Character Strings:

Returned length ≥ Character variable length Returned length < Character variable length

NAME= specifier Other specifier NAME= specifier Other specifier

The returned value is padded with blanks to fill thecharacter variable.

A message is outputindicating that the lengthof the character variable istoo short.

The character variable isset to all blanks.

A message is outputindicating that the lengthof the character variable istoo short.

The returned valuetruncated to the length ofthe character variable.

The tables below list the returned values for the INQUIRE statement.

Returned values when a filename is specified

Inquiry specifier File exists File does not exist

Being used (*1) Not being used (*2)

EXIST True True False

OPENED True False False

NUMBER o -1 -1

NAMED True True True

NAME o o o

FLEN o 0 0

ACCESS o "UNDEFINED" "UNDEFINED"

(*8)

SEQUENTIAL o "UNKNOWN" "UNKNOWN"

(*8)

DIRECT o "UNKNOWN" "UNKNOWN"

(*8)

ACTION o "UNDEFINED" "UNDEFINED"

(*8)

READ o "UNKNOWN" "UNKNOWN"

(*8)

WRITE o "UNKNOWN" "UNKNOWN"

(*8)

READWRITE o "UNKNOWN" "UNKNOWN"

(*8)

- 128 -

Inquiry specifier File exists File does not exist

Being used (*1) Not being used (*2)

FORM o "UNDEFINED" "UNDEFINED"

(*8)

FORMATTED o "UNKNOWN" "UNKNOWN"

(*8)

UNFORMATTED o "UNKNOWN" "UNKNOWN"

(*8)

BINARY o "UNKNOWN" "UNKNOWN"

(*8)

RECL o

(*3)

0 0

NEXTREC o

(*4)

0 0

BLANK o

(*5)

"UNDEFINED" "UNDEFINED"

(*8)

PAD o

(*5)

"UNDEFINED"

(*7)

"UNDEFINED"

(*13)

POSITION o

(*6)


(*8)

DELIM o

(*5)


(*8)

BLOCKSIZE o

(*9)

0 0

CARRIAGECONTROL o

(*5)


(*8)

CONVERT o

(*10)

"UNKNOWN" "UNKNOWN"

(*8)

ASYNCHRONOUS o

(*11)

“UNDEFINED” “UNDEFINED”

DECIMAL o “UNDEFINED” “UNDEFINED”

ENCODING o “UNKNOWN” “UNKNOWN”

ROUND o “UNDEFINED” “UNDEFINED”

SIGN o “UNDEFINED” “UNDEFINED”

ID o 0 0

PENDING o FALSE FALSE

POS o

(*12)

0 0

SIZE o o -1

STREAM o “UNKNOWN” “UNKNOWN”

- 129 -

True or False: Four-byte logical value

o : A defined value is assigned to the variable

*1. This column indicates the processing when input/output statements are being executed for the relevant unit number.

*2. This column indicates the processing when input/output statements are not being executed for the relevant unit number.

*3. The value of the RECL= specifier in the OPEN statement is assigned to the variable. If the RECL= specifier in the OPEN statementis omitted, the value 2147483647 in Fortran95 and Fortran 2003 or the value 0 in FORTRAN66 and FORTRAN77 is assigned to thevariable.

*4. Set only for direct access; otherwise, zero.

*5. Set only for formatted input/output; otherwise, it is assigned the value UNDEFINED.

*6. Set only for sequential access; otherwise, it is assigned the value UNDEFINED.

*7. Set the value YES in Fortran95.

*8. Set blank characters in FORTRAN66 and FORTRAN77.

*9. Set only for sequential access; otherwise, zero.

*10. Set only for unformatted input/output; otherwise, it is assigned the value NATIVE.

*11. For a file or unit for which access other than unformatted sequential is specified, or when a special file has been connected, it isassigned the value NO.

*12. Set only for sequential access; otherwise, it is assigned the value 0.

*13. Set the value YES in Fortran95. Set blank characters in FORTRAN66 and FORTRAN77.

Note:

1. In FORTRAN66 and FORTRAN77, the return value of the character type is returned by the small letter.

2. The undefined value is the blank character or the value 0. The blank character is assigned to the character variable. 0 is assigned tothe integer variable.

3. If the Fortran system cannot define a meaningful value, -1, 0, blanks, "UNKNOWN" or "UNDEFINED" is assigned to the variable.

Returned values when unit is specified.

Inquiry specifier Unit number within the range (*1) Unit specifier number

outside the range (*2)File connected (*3) File not connected(*4)Being used

(*5)Not being used (*6)

EXIST True True True False

OPENED True True False False

NUMBER o -1 -1 -1

NAMED True

(*7)

False False False

NAME o

(*8)

" " " " " "

FLEN o 0 0 0

ACCESS o "UNDEFINED" "UNDEFINED" "UNDEFINED"

(*14)

SEQUENTIAL o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

DIRECT o "UNKNOWN" "UNKNOWN" "UNKNOWN"

- 130 -

Inquiry specifier Unit number within the range (*1) Unit specifier numberoutside the range (*2)File connected (*3) File not connected

(*4)Being used(*5)

Not being used (*6)

(*14)

ACTION o "UNDEFINED" "UNDEFINED" "UNDEFINED"

(*14)

READ o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

WRITE o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

READWRITE o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

FORM o "UNDEFINED" "UNDEFINED" "UNDEFINED"

(*14)

FORMATTED o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

UNFORMATTED o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

BINARY o "UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

RECL o

(*9)

0 0 0

(*14)

NEXTREC o

(*10)

0 0 0

(*14)

BLANK o

(*11)

"UNDEFINED" "UNDEFINED" "UNDEFINED"

(*14)

PAD o

(*11)

"UNDEFINED"

(*19)

"UNDEFINED"

(*19)

"UNDEFINED"

(*20)

POSITION o

(*12)


(*14)

DELIM o

(*11)


(*14)

BLOCKSIZE o

(*13)

0 0 0

CARRIAGECONTROL o

(*15)


(*14)

CONVERT o

(*16)

"UNKNOWN" "UNKNOWN" "UNKNOWN"

(*14)

ASYNCHRONOUS o(*17)

“UNDEFINED” “UNDEFINED” “UNDEFINED”

- 131 -

Inquiry specifier Unit number within the range (*1) Unit specifier numberoutside the range (*2)File connected (*3) File not connected

(*4)Being used(*5)

Not being used (*6)

DECIMAL o “UNDEFINED” “UNDEFINED” “UNDEFINED”

ENCODING o “UNKNOWN” “UNKNOWN” “UNKNOWN”

ROUND o “UNDEFINED” “UNDEFINED” “UNDEFINED”

SIGN o “UNDEFINED” “UNDEFINED” “UNDEFINED”

ID o 0 0 0

PENDING o FALSE FALSE FALSE

POS o(*18)

0 0 0

SIZE o o -1 -1

STREAM o “UNKNOWN” “UNKNOWN” “UNKNOWN”

True or False: Four-byte logical value

o : A defined value is assigned to the variable

" ": Indicates blanks.

*1. The unit number range is 0 to 2147483647.

*2. Outside the unit number means less than 0 or greater than 2147483647.

*3. This column indicates the processing when a unit number and file are connected.

*4. This column indicates the processing when a unit number and file are not connected.

*5. This column indicates the processing when input/output statements are being executed for the relevant unit number.

*6. This column indicates the processing when input/output statements are not being executed for the relevant unit number.

*7. For an unnamed file, false is set.

*8. For an unnamed file, a blank is set.

*9. The value of the RECL= specifier in the OPEN statement is assigned to the variable. If the RECL= specifier in the OPEN statementis omitted, the value 2147483647 in Fortran95 and Fortran 2003 or the value 0 in FORTRAN66 and FORTRAN77 is assigned to thevariable.

*10. Set only for direct access; otherwise, zero.

*11. Set only for formatted input/output; otherwise, it is assigned the value UNDEFINED.

*12. Set only for sequential access; otherwise, it is assigned the value UNDEFINED.

*13. Set only for sequential access; otherwise, zero.

*14. Set blank characters in FORTRAN66 and FORTRAN77.

*15. Set only for formatted sequential access, otherwise, UNDEFINED.

*16. Set only for unformatted input/output; otherwise, it is assigned the value NATIVE.

*17. For a file or unit for which access other than unformatted sequential is specified, or when a special file has been connected, it isassigned the value NO.

*18. Set only for sequential access; otherwise, it is assigned the value 0.

*19. Set the value YES in Fortran95.

*20. Set the value YES in Fortran95. Set blank characters in FORTRAN66 and FORTRAN77.

Note:

- 132 -

1. In FORTRAN66 and FORTRAN77, the return value of the character type is returned by the small letter.

2. The undefined value is the blank character or the value 0. The blank character is assigned to the character variable. 0 is assigned tothe integer variable.

3. If the Fortran system cannot define a meaningful value, -1, 0, blanks, "UNKNOWN" or "UNDEFINED" is assigned to the variable.

7.6.2 Difference in Language Version of the INQUIRE statementSee Section 7.6.1 INQUIRE Statement Parameters for details of the value returned by the INQUIRE statement.

7.6.2.1 Return value of ACTION= specifierIn FORTRAN66 and FORTRAN77, the variable in the ACTION= specifier is assigned the value BOTH if the file is connected for bothinput and output.

In Fortran95 and Fortran 2003, it is assigned the value READWRITE.

7.6.2.2 Char-constant returned by inquiry specifiers of INQUIRE statementIn Fortran66 and FORTRAN77, char-constant returned by an inquiry specifier except NAME= specifier is returned by the small letter.However, when run-time option -q is specified, it is returned by the capital letter.

In Fortran95 and Fortran 2003, it is returned by the capital letter.

Example:

CHARCTER(LEN=10) CHOPEN(10,ACCESS='SEQUENTIAL')INQUIRE(10,ACCESS=CH)

In FORTRAN66 and FORTRAN77, CH is assigned the value sequential.

In Fortran95 and Fortran 2003, CH is assigned the value SEQUENTIAL.

7.6.2.3 Return value of PAD= specifierIn FORTRAN66, FORTRAN77 and Fortran2003, the variable of PAD= specifier is assigned the value undefined or the blank characterif the file is not being used.

In Fortran95, it is assigned the value YES.

7.6.2.4 Return value of the inquire by file of INQUIREstatementIn FORTRAN66 and FORTRAN77, the variable of the following specifiers is assigned the blank character if the file does not exist.

In Fortran95 and Fortran 2003, it is assigned the value UNDEFINED or UNKNOWN.

Specifiers are ACCESS=, SEQUENTIAL=, DIRECT=, ACTION=, READ=, WRITE=, READWRITE=, FORM=, FORMATTED=,

UNFORMATTED=, BINARY=, BLANK=, POSITION=, DELIM=, ASYNCHRONOUS=, DECIMAL=, ROUND=, SIGN= andSTREAM=.

7.6.2.5 Return value of the inquire by unit of INQUIRE statementIn FORTRAN66 and FORTRAN77, the variable of the following specifiers is assigned the value the blank character if UNIT= specifieris out of range from 0 to 2147483647.

In Fortran95 and Fortran 2003, it is assigned UNDEFINED or UNKNOWN.

Specifiers are ACCESS=, SEQUENTIAL=, DIRECT=, ACTION=, READ=, WRITE=, READWRITE=, FORM=, FORMATTED=,

UNFORMATTED=, BINARY=, BLANK=, POSITION=, DELIM=, ASYNCHRONOUS=, DECIMAL=, ROUND=, SIGN= andSTREAM=.

- 133 -

7.7 Data Transfer Input/output StatementsNot every input/output statement may be used with all input/output devices. For example, a terminal device must not be accessed usingunformatted I/O.

Input/output statement type Terminal Direct access storage device

Formatted sequential (includingnamelist and list-directed)

Allowed

Unformatted sequential

Not allowed AllowedDirect

Stream

7.7.1 Control Information for Input/output Data Transfer StatementsThis section describes the control information (configuration elements) used in data transfer input/output statements.

The control specifiers are the UNIT=, FMT=, IOSTAT=, ERR=, END=, EOR=, and IOMSG= specifiers.

7.7.1.1 Input/output UNIT SpecifierWhen the user specifies an asterisk as the unit number or omits the UNIT= specifier, the system uses a unit based on a runtime option. Ifno runtime option is specified, the system uses the default specification. The table below lists the standard unit numbers used in theconditions just described.

Standard unit numbers:

Input/output statement Runtime option Standard value

Diagnostic message output by the Fortransystem

Value of u_no specified by runtimeoption -mu_no

0

READ statement with * specified.

READ statement with the unit specifieromitted.

Value of u_no specified by runtimeoption -ru_no

5

WRITE statement with * specified.

PRINT statement.

Value of u_no specified by runtimeoption -pu_no

6

Note:

Values specified using runtime options have priority over the default values.

7.7.1.2 Format SpecifierData is edited as real, complex, integer, logical, character, binary, octal, or hexadecimal according to an edit descriptor. For example,specifying the I edit descriptor for type real output edits data to produce integer data. The table below lists the edit descriptors and typesof input/output.

Table 7.1 Edit descriptors and types of input/output

Types of input Edit descriptor output fields

L I F,E,D,Q,EN,ES

A G B Z O

One-byte logical

Two-byte logical

Four-byte logical

Eight-byte logical

o + + o o o o o

- 134 -

Types of input Edit descriptor output fields

L I F,E,D,Q,EN,ES

A G B Z O

One-byte integer

Two-byte integer

Four-byte integer

Eight-byte integer

+ o + o o o o o

Default real



+ + o o o o o o

Default complex


Quadruple-precisioncomplex

+ + o o o o o o

Character x x x o o o o o

o: Combination is allowed.

+: Conflicting combination; a warning message is output and editing is performed according to the edit descriptor.

x: Combination is not allowed. A message is output, and processing terminates.

7.7.1.3 IOSTAT= SpecifierWhen an data transfer input/output statement is executed, the IOSTAT= specifier is set indicating one of the following:

- An error condition occurs.

- An end-of-file condition occurs.

- An end-of-record condition occurs.

- Neither an error condition, an end-of-file condition nor an end-of-record condition occurred.

The tables below list the value assigned to the IOSTAT= specifier after an error condition, an end-of-file condition, or an end-of-recordcondition occurs. The IOSTAT= specifier is set to zero if neither an error condition, an end-of-file condition, nor an end-of-record conditionoccurs. If an end-of-file condition or an end-of-record condition occurs for an input statement containing an IOSTAT= specifier, adiagnostic message is not output. The standard system and user-defined corrections for errors do not occur. See Section 8.1 ErrorProcessing for details.

Error, end-of-file and end-of-record condition

Input/output error

conditionsError cases Corresponding input/output

statementValue of IOSTAT=

specifier

Error condition(Unrecoverableerror)

An unrecoverable error occurredduring file input/output.

OPEN statement

CLOSE statement

Data transfer statements

File positioning statements

FLUSH statement

1

Error condition(exceptunrecoverable error)

An I/O error was detected.

OPEN statement

CLOSE statement

Data transfer statements

Error number

- 135 -

Input/output errorconditions

Error cases Corresponding input/outputstatement

Value of IOSTAT=specifier

File positioning statements

FLUSH statement

End-of-filecondition

An end-of-file record was detected.

Sequential READ

List-directed READ

Namelist READ

Stream READ

-1

End-of-filecondition

A Fortran record exceeded the endof the internal file.

Internal file READ -1

End-of-recordcondition

The length of input list exceeds thelength of a Fortran record.

Nonadvancing formatted sequentialinput

-2

7.7.1.4 Error BranchIf an error condition occurs during the execution of a data transfer input/output statement, execution continues with the statement specifiedin the ERR= specifier. If the input/output statement also contains an IOSTAT= specifier, the value of its error number is assigned to thevariable. See Section 7.7.1.3 IOSTAT= Specifier for details.

If an error condition occurs for an input/output statement that specifies the ERR= specifier, a diagnostic message is not output, and thestandard system and user-defined error correction do not occur.

7.7.1.5 End-of-File BranchIf an end-of-file condition occurs and no error condition occurs during the execution of a data transfer input/output statement, executioncontinues with the statement specified in the END= specifier. If the input/output statement also contains an IOSTAT= specifier, the value-1 is assigned to the variable.

If an end-of-file condition occurs for an input statement that specifies the END= specifier, a diagnostic message is not output, and thestandard system and user-defined error correction do not occur.

7.7.1.6 End-of-Record BranchIf an end-of-record condition occurs and no error condition occurs during the execution of a data transfer input/output statement, executioncontinues with the statement specified in the EOR= specifier. If the input/output statement also contains an IOSTAT= specifier, the value-2 is assigned to the variable.

If an end-of-record condition occurs for an input statement that specifies the EOR= specifier, a diagnostic message is not output, and thestandard system and user-defined error correction do not occur.

7.7.1.7 IOMSG= SpecifierThe IOMSG= specifier sets an error message when an error condition, end-of-file condition, or end-of-record condition occurs.

Note, however, that if an IOSTATE= specifier or an ERR= specifier is not specified simultaneously with the IOMSG= specifier, a messageis not set. Further, if an error condition, end-of-file condition, or end-of-record condition does not occur, the value specified in the IOMSG=specifier is not changed.

7.7.2 Sequential Access Input/output StatementsSequential input/output statements read and write formatted and unformatted Fortran records from sequential files. Formatted sequentialinput/output statements are used to read and write formatted Fortran records. Unformatted sequential input/output statements are used toread and write unformatted Fortran records.

- 136 -

7.7.2.1 Formatted Sequential Access Input/output StatementsFormatted sequential input/output statements sequentially read and write characters based on a format specification. Generally, thesestatements are used when the data must be displayed.

Formatted sequential READ statements read data from the current Fortran record based on a format. When an input list is specified, theFortran record is edited into internal representations based on edit descriptors in the format.

Formatted sequential WRITE and PRINT statements write Fortran records to the current record based on a format. When an output list isspecified, each item is edited based on the edit descriptors in the format. If the $ edit descriptor is specified, a line-feed character is notwritten after the edited data.

Example: A formatted sequential input/output statement used to read and write data

REAL A,B INTEGER I,J A=1.5 !A is assigned the hexadecimal value 3fc00000 I=100 !I is assigned the hexadecimal value 00000064 WRITE(1,10) I,A !A Fortran record becomes the hexadecimal ! value 2031303020312e35 REWIND 1 !Position the file at the beginning. READ(1,10) J,B !J is assigned the hexadecimal value 00000064. ! B is assigned the hexadecimal value 3fc00000. 10 FORMAT(1X,I3,1X,F3.1) STOP END

7.7.2.2 Unformatted Sequential Access Input/output StatementsUnformatted sequential input/output statements read and write Fortran records consisting of internal representations. Generally, thesestatements are used when the data need not be displayed.

If unformatted sequential input/output statements are executed for devices (such as terminals) that process character codes, the results arenot predictable.

Files written with variable-length Fortran records by Fujitsu Fortran implementation can be accessed by an unformatted input statement.

Unformatted sequential READ statements read the current record. When an input list is specified, the data is stored unchanged into eachitem in the input list in the order specified. An error occurs if the number of bytes in the record is less than in the input list items. If thenumber of bytes in the record is greater, the excess data is ignored. If an input list is not specified, the record is skipped.

Unformatted sequential WRITE statements write the output list items as a Fortran record in the order they are specified. The statementwrites the record in the current file position. If an output list is not specified, a Fortran record of length zero is written.

The length of the Fortran record is the sum of the bytes in the output list items.

Example: An unformatted sequential input/output statement used to read and write data

REAL A,BINTEGER I,JA=1.5 !A is assigned the hexadecimal value 3fc00000I=100 !I is assigned the hexadecimal value 00000064WRITE(1) A,I !A Fortran record becomes the hexadecimal value ! 3fc0000000000064REWIND 1 !Position the file at the beginning.READ(1) B,J !B is assigned the hexadecimal value 3fc00000 ! J is assigned the hexadecimal value 00000064STOPEND

7.7.3 Direct Access Input/output StatementsDirect input/output statements are used to read and write formatted or unformatted Fortran records to or form the record of the file indicatedby the REC= specifier.

- 137 -

An OPEN statement must be executed before any direct input/output statements may be executed.

When TOTALREC is used in an OPEN statement, reading and writing can be executed only for the number of records specified. Sequentialinput/output statements must not be used to add records to direct files.

7.7.3.1 Direct Access READ StatementA direct access READ statement read Fortran records which is uniquely identified by the record number.

During formatted data transfer, data are transferred with editing between the file and the entities specified by the input list.

During unformatted data transfer, data are transferred without editing between current record and the entities specified by the input list.The number of values required by the input list shall be less than or equal to the number of values in the record.

Data are not transferred if the input list does not exist.

7.7.3.2 Direct Access WRITE StatementA direct access WRITE statement writes Fortran records which is uniquely identified by the record number.

During formatted data transfer, data are transferred with editing between the file and the entities specified by the output list. Blank valuesare transferred if the output list does not exist.

During unformatted data transfer, data are transferred without editing between current record and the entities specified by the output list.The output list shall not specify more values than can fit into the record. If the values specified by the output list do not fill the record, theremainder of the record is undefined. Zero values are transferred if the output list does not exist.

7.7.3.3 The Error of Data Transfer in a DIRECT Access Input/outputThe execution of Fortran program is different by the compiler option -X if the output list specify more values than can fit into the recordand the number of values required by the input list are more than the number of values in the record.

In FORTRAN66 and FORTRAN77, the program ends normally.

In Fortran 95 and Fortran 2003, a diagnostic message of w level is generated and the multiple records are written or read.

Example:

program di.f

OPEN(10,FILE='XX.DATA',FORM='UNFORMATTED',ACCESS='DIRECT',RECL=4)WRITE(10,REC=1)1.0_8INQUIRE(10,NEXTREC=i)PRINT*,'NEXTREC=',iEND

result


$ ./a.out NEXTREC=3$


$ ./a.outjwe1162i-w line 2 On output to a file connected for unformatted direct access, the output list must not specify more values than can fit into the Fortran record (unit=10). error occurs at MAIN__ line 2 loc 0000000000400be6 offset 0000000000000036 MAIN__ at loc 0000000000400bb0 called from o.s.taken to (standard) corrective action, execution continuing. NEXTREC= 3error summary (Fortran)error number error level error count jwe1162i w 1

- 138 -

total error count = 1$

7.7.4 Namelist Input/output StatementsNamelist input/output statements transfer data between memory and sequential formatted records in an internal or external file. Thenamelist names and their associated variable and array names used for data transfer must have been declared previously in a NAMELISTstatement.

7.7.4.1 Namelist READ StatementsIt is necessary for the type of the item in the namelist to match the type of the data in the record. The correspondence for type of elementand type of constant is indicated in the following table.

Type of namelist group

objectType of constant

Logical Integer Real Complex Character Hexadecimal

One-byte logical

Two-byte logical

Four-byte logical

Eight-byte logical

o +(*1)

+(*1)

x +(*2)

+(*7)(*8)

One-byte integer

Two-byte integer

Four-byte integer

Eight-byte integer

+(*3)

o +(*3)

x +(*2)

o(*8)

Default real



+(*4)

o(*5)

o x +(*2)

+(*7)(*8)

Default complex



+(*4)(*6)(*9)

+(*5)(*6)(*9)

+(*6)(*9)

o +(*2)

+(*7)(*8)

Characterx x x x o x

(*7)(*8)

o: Operation is normal.

+: Combination is not allowed, but operation continues; interprets to revised for element type.

x: Input is terminated because the combination is not allowed.

*1. Logical editing is performed.

*2. Character editing is performed.

*3. Integer editing is performed.

*4. Real editing is performed.

*5. Real editing is performed. The decimal point is assumed to be to the right of the constant.

*6. The edit result is entered in the real part. Zero is entered in the imaginary part.

*7. In FORTRAN66 and FORTRAN77, the operation for hexadecimal types is normal.

- 139 -

*8. In Fortran 95 and Fortran 2003, hexadecimal is assumed to be character. Therefore, if the type of element is character, the operationis normal. If the type of element is not character, the combination is not allowed but operation continues. See Section "2.2 CompilerOptions" for compiler options.

*9. In Fortran 95 and Fortran 2003, a diagnostic message of w level is generated.

Note:

For input processing, editing is based on the type of namelist element. For hexadecimal or character types, editing is based on the typeof constant in the file.

a. Processing of Character Constants for Array Name.The processing of character constants for array names depends on whether the type of the array is character or not. If the type ofthe array is character, one character constant corresponds to one array element. Even if the character constant exceeds the arrayelement length, there is no overflow to the next array element.

Example: Character constant processing for elements of type character Fortran program:

CHARACTER(4) HA(5)NAMELIST/NAMEH/HAREAD(1,NML=NAMEH) !READ statement for namelist NAMEH

If the record to be read from the file connected to unit 1 is

&NAMEH HA=2*'NAMELIST READ STATEMENT' /

then the elements of HA will have the following values:

If the type of the array is not character, the character constant corresponding to the first array element is entered for the entire array.That is, if the length of the character constant exceeds one element, the character constant is stored in succeeding elements. If thelength of the character constant is shorter than the entire array, the remaining elements are unchanged. If input ends within anelement, the remaining part of the element is padded with blanks.

For the second and subsequent array elements, each character constant corresponds to a single element.

Even if the length of the character constant exceeds the array element length, no overflow occurs.

Example: The processing of character constants for elements of other than type character

Fortran program:

INTEGER IA(5)NAMELIST/NAMEI/IAREAD(1,NML=NAMEI) !READ statement for namelist NAMEI

The namelist record for unit number 1 is set as follows:

&NAMEI IA=2*'NAMELIST READ STATEMENT' /

The following is the result in character notation:

- 140 -

b. Input Editing OperationThe format of the namelist READ statement is determined based on the element type, constant type and constant size, as follows:

- If the type of constant and type of element are other than character, editing is performed according to the element type. The editdescriptor is Gw. The letter w indicates the size (between delimiters) of the constant to be entered.

- If the constant type is character, character editing is performed. The edit descriptor is Aw. The letter w indicates the number ofcharacters.

- For noncharacter constants with elements of type character, character editing is performed. The edit descriptor is Aw.

7.7.4.1.1 Namelist data in Namelist input

The namelist data in the namelist input has one of following formats:

&name [v=c[,c]...[,v=c[,c]...]...] /

&name [v=c[,c]...[,v=c[,c]...]...] &end

name: The namelist-group name specified in the corresponding NAMELIST input/output statement.

v: The variable name or corresponding array element, array section, or character substring specified in the NAMELIST statementcorresponding to the name.

c: NULL or one of the following forms:

const

r*const

r*

const: A literal constant other than a Hollerith, binary, octal, or hexadecimal constant.

r: A repeat specification. The argument is an unsigned integer constant.

r*const: It is equivalent to the sequential appearance of the const r times.

r*: It is equivalent to the sequential appearance of NULL r times.

If the input item is a character type and all following conditions are satisfied, the character constant need not be enclosed with apostrophesor quotation marks.

- A blank, comma, slash, or ampersand, which is considered to be a value separator, is not included in the character constant.

- A semicolon, which is considered to be a value separator, is not included in the character constant, when the decimal point symbol isa comma.

- Left parentheses, right parentheses, and percent sign are not included in the character constant.

- The character constant does not extend over more than one Fortran record.

- The first non-blank character is a character other than an apostrophe or quotation mark.

- 141 -

- The form of the input item is not r*const.

When a character constant does not end with an delimiter character (apostrophe or quotation mark), it is assumed to end with the firstblank, comma, or slash. Moreover, an apostrophe or quotation mark in the character string is not treated as two consecutive apostrophesor two consecutive quotation marks.

7.7.4.2 Namelist WRITE StatementsNamelist WRITE statements are used for output editing. These statements write a record based on the types of variables and arrays specifiedin the namelist. The output format for variable names consists of a variable name, equal sign, and edit result, in that order. A commaseparates constants. The output format for array names consists of an array name, equal sign, and edit result for each array element, in thatorder. A comma separates the edit results.

The output format for a namelist group object is as follows:

Type of namelist group object Output format (Edit descriptor)

One-byte logical G1,1H, (*1)

Two-byte logical G1,1H, (*1)

Four-byte logical G1,1H, (*1)

Eight-byte logical G1,1H, (*1)

One-byte integer G4,1H, (*1)

Two-byte integer G6,1H, (*1)

Four-byte integer G11,1H, (*1)

Eight-byte integer G20,1H, (*1)

Default real1P,E15.8,1H, (*1)

0P,G16.9,1H, (*1)

Double-precision real1P,E22.15,1H, (*1)

0P,G23.16,1H, (*1)

Quadruple-precision real1P,E43.34E4,1H, (*1)

0P,G44.35E4,1H, (*1)

Default complex1H(,1P,E15.8,1H,,E15.8,2H), (*2)

1H(,0P,G16.9,1H,,G16.9,2H), (*2)

Double-precision complex1H(,1P,E22.15,1H,,E22.15,2H), (*2)

1H(,0P,G23.16,1H,,G23.16,2H), (*2)

Quadruple-precision complex1H(,1P,E43.34E4,1H,,E43.34E4,2H), (*2)

1H(,0P,G44.35E4,1H,,G44.35E4,2H), (*2)

n-byte character

1H',character-contents,2H', (*3)

1H",character-contents,2H", (*4)

character-contents,1H, (*5)

*1. If an &end or a / comes immediately after the value generated by the G or E edit descriptor, 1H, is omitted.

*2. If an &end or a / comes immediately after the value generated by the G or E edit descriptor, 1H) is generated instead of 2H),.

*3. If an &end or a / comes immediately after the value delimited by quotes, 1H' is generated instead of 2H',.

*4. If an &end or a / comes immediately after the value delimited by apostrophes, 1H" is generated instead of 2H",.

*5. If an &end or a / comes immediately after the nondelimited character constants, 1H, is omitted.

Notes:

- 142 -

1. For types other than character, meaningless blanks in the edit result are deleted before output. The actual output length can thereforebe smaller than the field width of the edit descriptor.

2. Whether 1P,Ew1.d1 or 0P,Gw2.d2 is used for type real or complex depends on the absolute value of the data. 0P,Gw2.d2 is used ifthe absolute value of x is 0 or 0.1 <= x < 10**d2. Otherwise, 1P,Ew1.d1 is used.

7.7.4.3 NAMELIST Input/output Version Differencesa. Character type output

In FORTRAN66 and FORTRAN77, character constants are written using apostrophes as delimiters. In Fortran 95 and Fortran 2003,DELIM= can be specified as the delimiter of character constants in an OPEN statement. If DELIM= specifier is omitted in an OPENstatement or an OPEN statement is not executed, the DELIM= specifier is treated as NONE. In this case, character constants arewritten without delimiters.

Example:

CHARACTER(LEN=5)::CH='XXXXX'NAMELIST /X/CHOPEN(10)WRITE(10,NML=X)

In FORTRAN66 and FORTRAN77, the output is as follows:

&XCH='XXXXX'&end

In Fortran 95 and Fortran 2003, the output is as follows:

&X CH=XXXXX/

b. Format for terminating the namelist output statementIn FORTRAN66 and FORTRAN77, &end is written to terminate the namelist output statement. InFortran 95 and Fortran 2003, / is written.

Example:

INTEGER::IH=123NAMELIST /X/IHWRITE(10,NML=X)


&XIH=123&end


&X IH=123/

c. Blanks in namelist input/outputFORTRAN66 and FORTRAN77 treat blanks as part of the input data. Fortran 95 and Fortran2003 treat a blank as a delimiter.However, blanks between delimiters are treated as input data.

Example:

INTEGER,DIMENSION(3) :: IHNAMELIST /X/IHREAD(*,NML=X)

Assume that the input data is IH=1_2_3, (_ indicates a blank).

FORTRAN66 and FORTRAN77 treat this data as follows:

- 143 -

IH(1)=10203IH(2)=0IH(3)=0

Fortran 95 and Fortran 2003 treat this data as follows:

IH(1)=1IH(2)=2IH(3)=3

d. Hexadecimal constants specified as namelist input dataFORTRAN66 and FORTRAN77 treat hexadecimal constants specified as namelist input data as hexadecimal constants. Fortran 95and Fortran2003 treat these constants as character constants.

Example:

CHARACTER(LEN=5) CHNAMELIST /X/CHREAD(*,NML=X)

Assume that the input data is CH=Z3132333435,.


CH='12345'


CH='Z3132'

e. Hollerith constants specified as namelist input data.FORTRAN66 and FORTRAN77 treat Hollerith constants specified as namelist input data as Hollerith constants. Fortran 95 andFortran 2003 treat Hollerith constants specified as namelist input data as character constants.

Example:

CHARACTER(LEN=5) CHNAMELIST /X/CHREAD(*,NML=X)

Assume that the input data is CH=3H123,.


CH='123 '


CH='3H123'

f. Output of the first byte of the second and subsequent records when Fortran records contain character constants.In FORTRAN66 and FORTRAN77, a blank is written in the first byte of the second record. In Fortran 95 and Fortran 2003, thefirst byte contains a value.

Example:

CHARACTER(LEN=1000) CHNAMELIST /X/CHWRITE(10,NML=X)

In FORTRAN66 and FORTRAN77, a blank is written as follows:

CH='XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX <- the first record*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX' <- the second record^blank is written

- 144 -

In Fortran 95 and Fortran 2003, a value is written as follows:

CH='XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX <- the first recordXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX' <- the second record^value is written

g. Input data with the r* form or r*const form.In FORTRAN66 and FORTRAN77, when a r* form appears in input data, it does not mean 'repeat the null value r times'. For thoseversions, it means the repeat is applied to the following input data. If the r* constitutes the last characters of a Fortran record, therepeat is applied to the first input data of the next Fortran record. In Fortran 95 and Fortran 2003, r* means 'repeat the null value rtimes'.

Example:

INTEGER::I(4)=(/1,2,3,4/)NAMELIST /X/IREAD(*,NML=X)

Assume that the input data is I=2*_30,40, (_ indicates a blank).


I(1)=30I(2)=30I(3)=40I(4)=4


I(1)=1I(2)=2I(3)=30I(4)=40

h. The output of a group name or an object name of namelist output statement.In FORTRAN66 and FORTRAN77, a group name or an object name in the output is case sensitive.In Fortran 95 and Fortran 2003, it is in uppercase.

Example:

INTEGER::Iu=1NAMELIST /Xu/IuWRITE(10,NML=Xu)


&XuIu=1&end


&XU IU=1/

i. The output of a namelist Fortran record.In FORTRAN66 and FORTRAN77, the output is as follows.The first record : _&name (_ indicates a blank, name indicates namelist group object name)The second record : _v=c[,c]...[,v=c[,c]...]...(v indicates namelist group object list item, c indicates namelist output value)The last record : _&endIn Fortran 95 and Fortran 2003, the output is as follows.The first record : _&name v=c[,c]...[,v=c[,c]...].../

- 145 -

j. Output of real number zero.

The result of executing a real number type of variable group output with a value of 0 varies according to the language specificationlevel.

In the FORTRAN66, FORTRAN77, and Fortran 95 language specifications, output of real number zero is equivalent to the resultoutput when E edit is used.

In the Fortran 2003 language specifications, output of real number zero is equivalent to the result output when F edit is used.

Example:

REAL :: R = 0NAMELIST /DATA/RWRITE(*, NML=DATA)

When the Fortran program above is executed, in the Fortran 95 language specifications, &DATA R=0.00000000E+00/ is output.

When the Fortran program above is executed, in the Fortran 2003 language specifications, &DATA R=0.00000000/ is output.

7.7.5 List-Directed Input/output StatementsList-directed input/output statements transfer data between memory and formatted records based on the type of the data. Internal files maybe used for list-directed formatting.

7.7.5.1 List-Directed READ StatementsList-directed READ statements transfer records from external and internal files to elements in input lists.

The types of elements in input lists must correspond to the types of elements in data items as indicated by the following table.

Type of element Type of constant

Logical Integer Real Complex Character

One-byte logical

Two-byte logical

Four-byte logical

Eight-byte logical

o +(*1)

+(*1)

x +(*2)

One-byte integer

Two-byte integer

Four-byte integer

Eight-byte integer

+(*3)

o +(*3)

x +(*2)

Default real



+(*4)

o(*5)

o x +(*2)

Default complex



+(*4)(*6)(*7)

+(*5)(*6)(*7)

+(*6)(*7)

o +(*2)

Character +(*2)

+(*2)

+(*2)

+(*2)

o

o: Operation is normal.

+: Combination is not allowed, but operation continues; interprets to revised for element type.

x: Input is terminated because the combination is not allowed.

- 146 -

*1. Logical editing is performed.

*2. Character editing is performed.

*3. Integer editing is performed.

*4. Real editing is performed.

*5. Real editing is performed. The decimal point is assumed to be to the right of the constant.

*6. The edit result is set in the real part. Zero is set in the imaginary part.

*7. In Fortran95 and Fortran 2003, a diagnostic message of w level is generated.

a. Processing of Character Constants for Array NamesThe processing of character constants for array names depends on whether the element is type character.If the element is type character, input is performed so that one character constant corresponds to one array element. Even if thelength of the character constant exceeds the array element, there is no overflow to the next array element.If the element is not type character, the character constant corresponding to the first array element is entered for the entire array.That is, if the length of the character constant exceeds one element, the succeeding elements are also read. If the length of thecharacter constant does not exceed the length of the entire array, the remaining elements are unchanged. If input ends within anelement, the remaining part of the element is padded with blanks. For second and subsequent array elements, the character constant corresponds to each element. Even if the length of the characterconstant exceeds that of the array element, the character constant does not overflow to the next element.

Example: How to enter data of type character for array names

Fortran program:

CHARACTER(LEN=4),DIMENSION(5) :: CHAR5INTEGER,DIMENSION(5) :: I4 ...READ(5,*) CHAR5 ! List-directed READ statement for ! character type array elementsREAD(5,*) I4 ! List-directed READ statement for ! other than character type array elements

The following is the format of the list-directed record for unit number 5:

3*'ABCDEFGHIJKLMNOPQ'/3*'ABCDEFGHIJKLMNOPQ'/

The following is the result for CHAR5 and I4:

b. Input Editing OperationThe format used by a list-directed READ statement is determined according to the element type, constant type, and constant size,as follows:

- If the constant and element are other than type character, editing is performed according to the element type. The edit descriptoris Gw. The letter w indicates the size (between delimiters) of the constant to be entered.

- If the constant is type character, character editing is performed. The edit descriptor is Aw. The letter w indicates the size betweenapostrophes.

- 147 -

- If the constant is other than type character and the element is type character, character editing is performed. The edit descriptoris Aw.

7.7.5.2 List-Directed WRITE StatementsList-directed WRITE statements generate formatted list-directed records. The following edit descriptors are used for output with a list-directed WRITE statement.

Type of element Output format (edit descriptor)

One-byte logical G1,1H_ (*1)

Two-byte logical G1,1H_ (*1)

Four-byte logical G1,1H_ (*1)

Eight-byte logical

One-byte integer

G1,1H_ (*1)

G4,1H_ (*1)

Two-byte integer G6,1H_ (*1)

Four-byte integer G11,1H_ (*1)

Eight-byte integer G20,1H_ (*1)

Default real 1P,E15.8,1H_ or 0P,G16.9,1H_ (*1)

Double-precision real 1P,E22.15,1H_ or 0P,G23.16,1H_ (*1)

Quadruple-precision real 1P,E43.34E4,1H_ or 0P,G44.35E4,1H_ (*1)

Default complex 1H(,1P,E15.8,1H,,E15.8,2H)_ or (*2)

1H(,0P,G16.9,1H,,G16.9,2H)_ (*2)

Double-precision complex 1H(,1P,E22.15,1H,,E22.15,2H)_ or (*2)

1H(,0P,G23.16,1H,,G23.16,2H)_ (*2)

Quadruple-precision complex 1H(,1P,E43.34E4,1H,,E43.34E4,2H)_ or (*2)

1H(,0P,G44.35E4,1H,,G44.35E4,2H)_ (*2)

n-byte character Gn,1H_ or (*3)(*6)

1H',Gn,2H'_ or (*4)(*6)

1H",Gn,2H"_ (*5)(*6)

_: Indicates blank

*1. If the record terminates immediately after the value generated by the G or E edit descriptor, 1H_ is omitted.

*2. If the record terminates 1 byte after the value generated by the G or E edit descriptor, 1H) is generated instead of 2H)_.

*3. If the record terminates immediately after the nondelimited character constants generated by the G edit descriptor, 1H_ is omitted.If the compiler option -X9 is not in effect, the output format is Gn.

*4. If the record terminates 1 byte after the value delimited by quotes generated by the G or E edit descriptor, 1H' is generated insteadof 2H'_.

*5. If the record terminates 1 byte after the value delimited by apostrophes generated by the G or E edit descriptor, 1H" is generatedinstead of 2H"_.

*6. In Fortran95 and Fortran 2003, is not in effect, the last blank is not output.

Notes:

1. For types other than character, meaningless blanks in the edit result are deleted before output. The actual output length can thereforebe smaller than the field width of the edit descriptor.

2. Whether 1P,Ew1.d1 or 0P,Gw2.d2 is used for type real or complex depends on the absolute value of the data. 0P,Gw2.d2 is used ifthe absolute value of x is 0 or 0.1 ≤ x < 10**d2. Otherwise, 1P,Ew1.d1 is used.

- 148 -

Example: List-directed WRITE statement

Fortran program:

INTEGER :: YY=2002, MM=09, DD=10CHARACTER(15) :: CH='YEAR MONTH DATE'INTEGER,DIMENSION(2,3) :: TABLEWRITE(6,*) YY,CH(1:4) ,MM,CH(6:10) ,DD,CH(12:15)OPEN(6,DELIM='QUOTE')WRITE(6,*) YY,CH(1:4) ,MM,CH(6:10) ,DD,CH(12:15)OPEN(6,DELIM='APOSTROPHE')WRITE(6,*) YY,CH(1:4) ,MM,CH(6:10) ,DD,CH(12:15)TABLE = RESHAPE((/1,2,3,4,5,6/) ,(/2,3/))WRITE(6,*) ((TABLE(I,J) ,I=2,1,-1) ,J=3,1,-1)END

The output is as follows:

$ ./a.out 2002 YEAR 9 MONTH 10 DATE 2002 "YEAR" 9 "MONTH" 10 "DATE" 2002 'YEAR' 9 'MONTH' 10 'DATE' 6 5 4 3 2 1$

7.7.5.2.1 The Form of Nondelimited Character Constants Produced by List-Directed Output

In FORTRAN66 and FORTRAN77, the values are not separated.

In Fortran 95 and Fortran 2003, the values are separated by one blank.

Example:

CHARACTER(10)::CH=' 'WRITE(*,*)'Year',2002WRITE(CH,*)'Month',9WRITE(*,'(A10)')CHEND


$ ./a.out Year2002 Month9$


$ ./a.out Year 2002 Month 9$

7.7.5.3 List-Directed Input/output Version Differences

Output of real number zero

The result of executing a real number type of list output with a value of 0 varies according to the language specification level.

In the FORTRAN66, FORTRAN77, and Fortran 95 language specifications, output of real number zero is equivalent to the result outputwhen E edit is used.

In the Fortran 2003 language specifications, output of real number zero is equivalent to the result output when F edit is used.

- 149 -

Example:

REAL :: R = 0WRITE(*,*) R

When the Fortran program above is executed, in the Fortran 95 language specifications, 0.00000000E+00 is output.

When the Fortran program above is executed, in the Fortran 2003 language specifications, 0.00000000 is output.

7.7.6 Internal File Input/output StatementsInternal file input/output statements treat a character string as a file. See Section 7.7.2 Sequential Access Input/output Statements, 7.7.4Namelist Input/output Statements, and 7.7.5 List-Directed Input/output Statements for details.

7.7.7 Nonadvancing Input/output StatementsNonadvancing input/output lets you control when the next record will be processed, and a single Fortran record can be read and writtenas many times as necessary. The statement may position the file at a character position within the current record. You can then use aSIZE=, IOSTAT=, or EOR= specifier to control the Fortran program. The variable specified in the SIZE= specifier becomes defined withthe count of the characters transferred by data edit descriptors during execution of the current input statement. If an end of record conditionoccurs, the control may be processed by an IOSTAT variable or EOR= specifier.

Example: Nonadvancing input statements with a SIZE= and EOR= specifier

CHARACTER(LEN=20) NAME INTEGER NO ! 1-20 : NAME ! 21-28 : NO OPEN(10,FORM='FORMATTED') DO READ(10,FMT='(A20) ',ADVANCE='NO',EOR=10,SIZE=IS,END=20) NAME IF (IS < 20) CYCLE READ(10,FMT='(I8) ',ADVANCE='NO',SIZE=IS,END=20,IOSTAT=IO) NO IF (IS < 8) CYCLE GO TO 30 10 CYCLE 30 PRINT *,NAME,' ',NO,' ',IS END DO GO TO 40 20 PRINT *,'END OF DATA' 40 END

7.8 Combining Input/output StatementsWhen input/output statements are executed, processing depends on the type of input/output statement executed previously.

The combinations of input/output statements explained here use the same unit number.

Processing according to input/output statement order

Current I/O

Immediately preceding I/O

Formattedsequential

Unformattedsequential

Formatteddirect

Unformatteddirect

Formattedstream

Unformattedstream

READ WRITE READ WRITE READ WRITE READ WRITE READ WRITE READ WRITE

FormattedsequentialREAD

o x

(*1)

x x x x x x x x x x

FormattedsequentialWRITE

o

(*2)

o x x x x x x x x x x

- 150 -

Current I/O


Formattedsequential

Unformattedsequential

Formatteddirect

Unformatteddirect

Formattedstream

Unformattedstream

READ WRITE READ WRITE READ WRITE READ WRITE READ WRITE READ WRITE

UnformattedsequentialREAD

x x o x

(*1)

x x x x x x x x

UnformattedsequentialWRITE

x x o

(*2)

o x x x x x x x x

Formatteddirectaccess

x x x x o o x

(*5)

x

(*5)

x x x x

Unformatteddirectaccess

x x x x x

(*5)

x

(*5)

o o x x x x

FormattedstreamREAD

x x x x x x x x o o

(*6)

x x

FormattedstreamWRITE

x x x x x x x x o o x x

UnformattedstreamREAD

x x x x x x x x x x o o

(*6)

UnformattedstreamWRITE

x x x x x x x x x x o o

OPEN o o o o o o o o o o o o

CLOSE o o o o o o o o o o o o

FLUSH - o - o - o - o - o - o

WAIT o o o o o o o o o o o o

BACKSPACE o o

(*3)

o o

(*3)

x x x x o o x x

REWIND o o o o x x x x o o o o

ENDFILE o

(*4)

o o

(*4)

o x x x x o o o o

o: The operation is normal.

x: A diagnostic message is output. Correction processing is performed.

-: No operation is performed.

*1. The READ statement cannot execute, because the Fortran record that follows the one written by the immediately preceding WRITEstatement is undefined.

*2. The Fortran record that follows the one written by the current WRITE statement is undefined.

*3. The endfile record is generated before BACKSPACE is executed.

*4. The endfile record is generated immediately after the READ statement reads the Fortran record.

- 151 -

*5. In FORTRAN66 and FORTRAN77, the operation is normal.

*6. Execution of a stream input statements omitted POS= specifier after the execution of a stream output statements is not allowed.

Note:

Formatted READ and WRITE statements also include list-directed and namelist input/output statements.

Current I/O


OPEN CLOSE FLUSH WAIT BACKSPACE REWIND END EOF

(*1)

None

FormattedsequentialREAD

o o

(*6)

o o o o

(*2)

o

(*7)

x

(*7)

o

(*8)

FormattedsequentialWRITE

o o

(*6)

o o o o

(*2)

o

(*6)

x

(*11)

o

(*8)

UnformattedsequentialREAD

o o

(*6)

o o o o

(*2)

o

(*7)

x

(*7)

o

(*8)

UnformattedsequentialWRITE

o o

(*6)

o o o o

(*2)

o

(*6)

x

(*11)

o

(*8)

Formatteddirectaccess

o x o o x x x x x

(*3)

Unformatteddirectaccess

o x o o x x x x x

(*3)

Formattedstreamaccess

o x o o o o

(*2)

o o x

(*3)

Unformattedstreamaccess

o x o o x o

(*2)

o o x

(*3)

OPEN o o o o o o o o o

CLOSE o o o o o o o o o

FLUSH - - - - - - - - -

WAIT o o o o o o o o o

BACKSPACE

o o

(*6)

o o o - o

(*5)

o

(*4)

o

(*10)

REWIND o o

(*9)

o o o - o o -

ENDFILE o

(*5)

o

(*6)

o o o o

(*5)

-

(*7)

-

(*9)

o

(*5)

o: The operation is normal.

x: A diagnostic message is output. Correction processing is performed.

-: No operation is performed.

- 152 -

*1. EOF means the end-of-file condition occurs.

*2. A combination of formatted and unformatted sequential input/output statements before or after a REWIND statement is not allowed.

*3. OPEN statement shall be executed for old file.

*4. BACKSPACE is executed for the endfile record.

*5. An empty file is created.

*6. Input/output is executed for the endfile record. In Fortran 2003 and Fortran 95, a diagnostic message (jwe0111i-e) is output. InFORTRAN766 and FORTRAN77, a diagnostic message (jwe0111i-e) is output.

*7. Input/output is executed for the endfile record.

*8. Input/output is executed for files named 'fort. nn' (nn:unit number.)

*9. Error.

*10. BACKSPACE is executed for the endfile record.

*11. Input/output is executed for the endfile record. In Fortran 2003 and Fortran 95, a diagnostic message (jwe0115i-e) is output. InFORTRAN766 and FORTRAN77, a diagnostic message (jwe0115i-e) is output.

Note:

Formatted READ and WRITE statements also include list-directed and namelist input/output statements.

7.8.1 Combinations of Input/output Statements Not AllowedSome combinations of input/output statements are not allowed because of the access method or file status.

An input/output statement combination that is not allowed is considered an error. After the error is corrected, execution continues.

7.8.1.1 Formatted Sequential Access Input/output StatementsIf the current statement is a formatted sequential (including list-directed and namelist) data transfer statement, the immediately precedinginput/output statement must not be one of the following statements:

- Unformatted sequential access data transfer

- Direct access data transfer

- Stream access data transfer

If the current statement is READ, the immediately preceding input/output statement must not be WRITE, except the input/output is aterminal.

7.8.1.2 Unformatted Sequential Access Input/output StatementsIf the current input/output statement is an unformatted sequential data transfer statement, the immediately preceding input/output statementmust not be any of the following statements:

- Formatted sequential data transfer

- List-Directed

- Namelist



If the current input/output statement is READ, the immediately preceding input/output statement must not be a WRITE statement.

7.8.1.3 Direct Access Input/output StatementsIf the current input/output statement is a direct access data transfer, the immediately preceding input/output statement must not be any ofthe following statements:

- Sequential data transfer

- 153 -

- List-Directed data transfer

- Namelist data transfer


- File positioning statement

- CLOSE

7.8.1.4 Stream Access Input/output StatementsIf the current input/output statement is a stream access data transfer, the immediately preceding input/output statement must not be anyof the following statements:

- Sequential data transfer


- List-Directed data transfer

- Namelist data transfer


- CLOSE

7.8.1.5 File Positioning StatementsBy combining sequential input/output statements with file positioning statements, you can control file positioning. The file positioningstatements are BACKSPACE, ENDFILE, and REWIND.

When a file is connected for sequential access, the initial file position is immediately before the first record in the file unless POSITIONis specified in the OPEN statement.

If the current input/output statement is BACKSPACE, which are used for file positioning, the immediately preceding input/ outputstatement must not be either of the following statements:


- Unformatted stream access data transfer

If the current input/output statement is REWIND, which are used for file positioning, the immediately preceding input/

output statement must not be either of the following statements:


If the current input/output statement is ENDFILE, which are used for file positioning, the immediately preceding input/ output statementmust not be either of the following statements:


7.9 File Positioning StatementsBy combining sequential input/output statements with file positioning statements, you can control file positioning. The file positioningstatements are BACKSPACE, ENDFILE, and REWIND.

- BACKSPACE Statement (see Section 7.9.1 BACKSPACE Statement)

- ENDFILE Statement (see Section 7.9.2 ENDFILE Statement)

- REWIND Statement (see Section 7.9.3 REWIND Statement)

7.9.1 BACKSPACE StatementExecution of a BACKSPACE statement causes the file connected to the specified unit to be positioned before the current record if thereis a current record, or before the preceding record if there is no current record. If there is no current record and no preceding record, the

- 154 -

position of the file is not changed. If the preceding record is an endfile record, the file is positioned before the endfile record. If aBACKSPACE statement causes the implicit writing of an endfile record, the file is positioned before the record that precedes the endfilerecord.

Example 1: A BACKSPACE statement after a READ statement

Fortran program:

READ(1) IDATA ! (a) Reads Fortran record 2BACKSPACE 1 ! (b) Position is before record 2READ(1) NDATA ! (c) Reads Fortran record 2END

File position:

As a result, the same data is read into IDATA and NDATA.

Example 2: A BACKSPACE statement after a WRITE statement

Fortran program:

WRITE(1) IDATA ! (a) Writes Fortran record 2BACKSPACE 1 ! (b) Position is before record 2READ(1) NDATA ! (c) Reads Fortran record 2END

File position:

As a result, the data written by IDATA is read into NDATA.

7.9.2 ENDFILE StatementThe execution of an ENDFILE statement writes an endfile record as the next record of the file. The file is then positioned after the endfilerecord, which becomes the last record of the file.

- 155 -

Example 1: An ENDFILE statement after executing a READ statement

Fortran program:

READ(1)IDATA ! (a) Reads Fortran record 2.ENDFILE 1 ! (b) Writes the endfile record.

File position:

Example 2: An ENDFILE statement after executing a WRITE statement

Fortran program:

WRITE(1) IDATA ! (a) Writes Fortran record 2.ENDFILE 1 ! (b) Writes the endfile record.

File position:

Example 3:An ENDFILE statement after executing a BACKSPACE statement

Fortran program:

WRITE(1)IDATA ! (a) Write Fortran record 2.BACKSPACE 1 ! (b) Write the endfile record.ENDFILE 1 ! (c) Erase Fortran record 2.

File position:

- 156 -

Example 4: An ENDFILE statement after executing a REWIND statement

Fortran Program:

READ(1)IDATA ! (a) Reads Fortran record 2.REWIND 1 ! (b) The file position is set to the beginning !of the file.ENDFILE 1 ! (c) Deletes all Fortran records and write the !endfile record.

File position:

As a result, the file consists of only an endfile record.

- 157 -

7.9.3 REWIND StatementWhen the file is connected using sequential or stream access, the file position is set to immediately before the first record in the file.

7.10 FLUSH statementExecution of a FLUSH statement causes data written to an external file to be available to other processes, or causes data placed in anexternal file by means other than Fortran to be available to a READ statement.

Execution of a FLUSH statement for a file that is connected but does not exist has no effect on any file. And when an immediately precedingI/O is not a write statement, it has no effect on any file in the same way.

The UNIT=, IOSTAT= and ERR= specifiers can be specified for the FLUSH statement. See Section 7.7.1 Control Information for Input/output Data Transfer Statements for information on these specifiers.

7.11 Conversion between IBM370-Format Floating-Point Data andIEEE-Format Floating-Point Data on Input/output

This section describes the conversion between IBM370-format floating-point data and IEEE-format floating-point data.

7.11.1 Relationship between Floating-Point Data and Input/outputStatements

If the runtime option -C is specified, the following operations are executed.

Unformatted READ statements read IBM370-format floating-point data in the unformatted Fortran records by converting them into internalIEEE-format floating-point data.

Unformatted WRITE statements convert internal IEEE-format floating-point data of the output list items into IBM370-format floating-point data and write the unformatted Fortran records. See Section 3.3 Runtime Options for details on runtime options.

The figures below show how IBM370-format floating-point data is stored in memory. See Section 5.1.3 Real and Complex Type Data fordetails on how IEEE-format floating-point data is stored in memory.

Default Real (Kind 4)

The figure below shows the bit composition of sign (s), exponent (e), fraction (f).

Real Storage Format

Double-Precision (Kind 8)

The figure below shows the bit composition of sign (s), exponent (e), fraction (f).

Double-Precision Storage Format

The flow of IBM370-format floating-point data

- 158 -

The following input/output statements are allowed:

- Unformatted sequential READ/WRITE statement

- Unformatted direct READ/WRITE statement

Files written with variable-length Fortran records by Fujitsu Fortran implementation can be converted.

See Section 7.3.2 Unformatted Fortran Records for details on unformatted records in this Fortran system.

The following types of input/output list items are allowed:

- Default real

- Double-precision real

- Default complex

- Double-precision complex

However, when a type of default complex or double-precision complex is specified for the input/output list items, the real and imaginaryparts are converted, respectively, to types of default real and double-precision real. Moreover, conversion processing is not performedwhen types other than default real, double-precision real, default complex, and double-precision complex are specified for input/outputlist items. The degree of accuracy when the floating-point data is converted is as follows:

Type of input/

output list itemsRange of values READ: Convert

IBM370-format toIEEE-format

WRITE: ConvertIEEE-format toIBM370-format

Real Exponent IBM370-formats

From 10**(-78) to10**75

Convertible withinthe following ranges.

From 10**(-37) to10**38 (*1)

Convertible

IEEE-formats

From 10**(-37) to10**38

Fraction IBM370-formats

24 bits Convertible Often, theconversion

- 159 -

Type of input/output list items

Range of values READ: ConvertIBM370-format to

IEEE-format

WRITE: ConvertIEEE-format toIBM370-format

IEEE-formats

caused a loss ofthree bits or less.(*2)

24 bits (include ahidden bit)


Exponent IBM370-formats

From 10**(-78) to10**75

Convertible Convertiblewithin thefollowing ranges.

From 10**(-78)to 10**75 (*1)

IEEE-formats

From 10**(-307) to10**308

Fraction IBM370-formats

56 bits Often, theconversion caused aloss of three bits orless. (*3)

Convertible

IEEE-formats

53 bits (include ahidden bit)

*1: The conversion may cause an overflow or underflow, if the value is outside the range.

*2: The conversion may cause a loss of up to three bits, when up to three bits are added to the top of the fraction bit field dependingon the exponent value.

*3: The conversion may cause a loss of up to three bits when the number of bits that the fraction occupies is different from ranges.

Note:

The conversion always sets the least significant bit of the fraction to 1 in order to minimize the error when the conversion causes aloss of bits. A diagnostic message is output if runtime option -M is specified.

7.11.2 Error Processing in the Conversion of Floating-Point Data

Conversion Data types Error event Diagnosticmessage

Standard correctionvalue

Convert IEEE-format data toIBM370-formatdata:

WRITE

Default real Inf is detected jwe0145i-w Maximum value

NaN is detected jwe0145i-w Maximum value

Missing bit is detected jwe0147i-w

(*1)

-------

Double-precision real Exponent overflow isdetected

jwe0142i-w Maximum value

Exponent underflow isdetected

jwe0144i-w 0

Inf is detected jwe0146i-w Maximum value

NaN is detected jwe0146i-w Maximum value

Convert IBM370-format data toIEEE-format data:

READ

Default real Exponent overflow isdetected

jwe0141i-w Maximum value

Exponent underflow isdetected

jwe0143i-w 0

Double-precision real Missing bit is detected jwe0147i-w

(*1)

-------

*1: A diagnostic message is output if the runtime option -M is specified.

- 160 -

7.12 Conversion between Little Endian Data and Big Endian Dataon Input/output

This section describes the conversion between little endian data and big endian data.

7.12.1 Relationship between Little Endian Data and Input/output StatementsYou may specify the runtime option -T to read and write little endian data. If the runtime option -T is specified, the following operationsare executed. See Section 3.3 Runtime Options for information on the runtime option -T.

The flow of little endian data




- Unformatted stream READ/WRITE statement

Files written with variable-length Fortran records by Fujitsu Fortran implementation can be converted.

See Section 7.3.2 Unformatted Fortran Records for details on unformatted records in this Fortran system.

The following types of input/output list items are allowed:

- 1-byte integer

- 2-byte integer

- 4-byte integer

- 8-byte integer

- Default real

- 161 -

- Double-precision real

- Quadruple-Precision real

- Default complex

- Double-precision complex

- Quadruple-Precision complex

- 1-byte logical

- 2-byte logical

- 4-byte logical

- 8-byte logical

7.13 Standard Files

7.13.1 Standard Files and Open ModesThe table below shows the relationship between standard files and open modes.

Input/output file Specification method (terminal open

mode if no specification)Open mode

Standard input command < file-name r

command << eof-sym r

Standard output command > file-name w

command >> file-name a

Standard error command 2> file-name w

command 2>> file-name a

7.13.2 Restrictions on Input/output Statements for Standard FilesOnly sequential access data transfer and non-data-transfer input/output statements may be used for standard input, standard output, anderror output files. Standard input files are processed as ACTION='READ'.

Standard output and standard error output files are processed as ACTION='WRITE'.

Executing an OPEN statement with a FILE= specifier for a standard input, standard output, or standard error output unit terminates anyexisting connection and opens a new file. The unit number of the standard input, standard output, or standard error output file is processedin the same way as a regular file.

7.13.3 Relationship between Input/output Statements and Seekable andNonseekable Files

Seekable files are files for which seek operations can be performed. The differences between seekable and nonseekable files are as follows:

Seekable or nonseekable File type

Seekable file Regular file

Nonseekable file Block-type special file

Character-type special file

FIFO special file

Note:

- 162 -

File type refers to information obtained from stat or fstat system calls.

The input/output statements and edit descriptors listed here may be executed for seekable files. If they are executed for nonseekable files,a diagnostic message is output, and the statement is ignored.

- REWIND statement

- BACKSPACE statement

- ENDFILE statement (INPUT system)

- Direct access statement

- Stream access statement

- Unformatted sequential statement

- TL and T edit descriptors (only when the same operation as that using the TL edit descriptors is performed)

7.14 Restrictions Related to the ACTION= SpecifierInput/output statements with files that have been connected using the ACTION= specifier have the following restrictions.

Input/outputstatement

ACTION= specifier

READ WRITE READWRITE

READ o x o

WRITE x o o

PRINT x o o

REWIND o x o

BACKSPACE o x o

OPEN o o o

CLOSE o o o

ENDFILE -- o o

o: Normal operation.

x: An error message is generated. Respond accordingly.

--: No operation is performed.

7.15 Number of Significant DigitsThe number of significant digits for real type is defined in this system as follows:

Type Number of significant digits

Default real 9

Double-precision real 16

Quadruple-precision real 35

Default complex 9

Double-precision complex 16

Quadruple-precision complex 35

7.16 I/O BufferThe table below shows how to change the size of a sequential I/O buffer. The precedence order is summarized in the below table.

- 163 -

Operators Size of buffer Precedence order

BLOCKSIZE= Specifier

Environment variable (fuxxbf)

Runtime option (-g)

Default

Defined Value

Defined Value

Defined Value

8M bytes

High

.

.

Low

7.17 Edit Descriptors

7.17.1 Generalized Integer EditingIn FORTRAN66 and FORTRAN77, the Gw.d and Gw.dEe edit descriptors follow the rules for the Iw.m edit descriptor when used tospecify the input/output of integer data.

In Fortran 95 and Fortran 2003, the Gw.d and Gw.dEe edit descriptors follow the rules for the Iw edit descriptor.

Example:

WRITE(*,'(G10.5)')123

In FORTRAN66 and FORTRAN77, the output is -----00123 (- indicates a blank).

In Fortran 95, the output is -------123.

7.17.2 Exponent Character in Formatted OutputIn FORTRAN66 and FORTRAN77, the exponent character is output in lowercase for E, EN, ES, D, G, L, Q, and Z editing. If runtimeoption -q is specified, the character is output in uppercase.

In Fortran 95 and Fortran 2003, the character is output in uppercase whether runtime option -q is specified or not.

Example:

WRITE(*,FMT='(1H E15.5)')6.5432

In FORTRAN66 and FORTRAN77, the output is 0.65432e+01.

In Fortran 95, the output is 0.65432E+01.

7.17.3 Result of G EditingIn FORTRAN66 and FORTRAN77, a zero value is edited using the E edit descriptor. In Fortran 95 and Fortran 2003, a zero value isedited using the F edit descriptor.

7.17.4 Diagnostic Messages for the Character String Edit DescriptorIn FORTRAN66 and FORTRAN77, the system outputs a level w diagnostic message (jwe0159i-w) for a character string edit descriptorin a format used for input. The character string edit descriptor is replaced by input data.

In Fortran 95 and Fortran 2003, the system outputs a level e diagnostic message (jwe1181i-e). The input list is ignored.

7.17.5 Effect of the X Edit DescriptorIf the X edit descriptor (nX) is used for output, the result differs between FORTRAN66 and other language version.

In FORTRAN77, Fortran 95 and Fortran 2003, the record character position is advanced by n characters.

Blanks are not inserted.

In FORTRAN66, n blanks are inserted in the record.

This operation differs only if a left-positioning tab is used:

- 164 -

Example:

FORMAT (5H FFFF,I4,T2,4X,I2)

The Fortran program operates as follows:

FORTRAN77, Fortran 95 and Fortran 2003 program: -FFFFjjiiFORTRAN66 program: -----jjii

The two low-order digits of I4 are represented by ii, I2 is represented by jj, and - indicates a blank.

For FORTRAN77, Fortran 95 and Fortran 2003 programs, the Fortran record is first padded with blanks.

7.17.6 L Edit DescriptorIn FORTRAN66 or the FORTRAN77, if the type of the input item corresponding to the Lw type edit descriptor is not a logical constant,the input item is assigned the value FALSE.

In Fortran 95 and Fortran 2003, a diagnostic message (jwe1202i-w) is output.

Example:

Program a.f:

LOGICAL :: L=.TRUE.OPEN (10,FILE='a.file')READ (10,FMT='(L1)') LPRINT *,LEND

file a.file:

1


$ ./a.out f$


$ ./a.outjwe1202i-w line 3 Invalid LOGICAL input (unit=10). error occurs at MAIN__ line 3 loc 0000000000400b96 offset 0000000000000036 MAIN__ at loc 0000000000400b60 called from o.s.taken to (standard) corrective action, execution continuing. Terror summary (Fortran)error number error level error count jwe1202i w 1total error count = 1$

7.17.7 D, E, F, G, I, L, B, O, and Z Edit Descriptor without w or d IndicatorsThe D, E, F, G, I, L, B, O, or Z edit descriptor may omit the width (w) or the number of decimals (d).

For the Iw, Bw, Ow, Zw, Fw, Ew, Dw, Lw, and Gw edit descriptors, the Fortran system selects the field width when w is omitted.

The table below shows the selected width.

- 165 -

Editdescriptor

Type

I1 I2 I4 I8 R4 R8 R16 C8 C16 C32 L1 L2 L4 L8

B 9 17 33 65 33 65 129 33 65 129 9 17 33 65

O 4 7 12 23 12 23 44 12 23 44 4 7 12 23

Z 3 5 9 17 9 17 33 9 17 33 3 5 9 17

I 4 6 11 20 11 11 11 11 11 11 4 6 11 20

L 2 2 2 2 2 2 2 2 2 2 2 2 2 2

F 15 15 15 22 15 22 43 15 22 43 15 15 15 22

E 15 15 15 22 15 22 43 15 22 43 15 15 15 22

D 15 15 15 22 15 22 43 15 22 43 15 15 15 22

G 4 6 11 20 15 22 43 15 22 43 2 2 2 2

I1: 1 byte integer

I2: 2 byte integer

I4: 4 byte integer

I8: 8 byte integer

R4: default real

R8: double-precision real

R16: quadruple-precision real

C8: default complex

C16: double-precision complex

C32: quadruple-precision complex

L1: 1 byte logical

L2: 2 byte logical

L4: 4 byte logical

L8: 8 byte logical

7.17.8 Editing of Special Real-Type DataWhen using one of the E, D, Q, G, F, EN, or ES edit descriptors, Inf or NaN are output right-justified within width w of the field when E,D, Q, G, F, EN, and ES edit descriptors are specified in the format Xw.d in Fortran 2003. In FORTRAN66, FORTRAN77 and Fortran95,Inf or NaN are output left-justified.

If the output width has fewer characters than the field width w, it will be padded with spaces. -Inf or -NaN is output for a negative value.

Moreover, the sign can be edited with the S/SP/SS edit descriptors.

Character strings (Inf/-Inf, Infinity/-Infinity, NaN/-NaN, NaN()/-NaN()) can be used as the input data of a real number. It is zero or morealphanumeric characters parentheses of NaN(). The positive sign "+" can be omitted.

The table below shows the correspondence between strings in a file and the values set in the variables.

Correspondence between Strings and Values Set in Variables

Strings Default real Double-precision Quadruple-precisio

+NaN

or +NAN()

0x7fffffff 0x7fffffff ffffffff 0x7fffffffffffffff ffffffffffffffff

-NaN 0xffffffff 0xffffffff ffffffff 0xffffffffffffffff ffffffffffffffff

- 166 -

Strings Default real Double-precision Quadruple-precisio

or -NAN()

+Inf

or+Infinity

0x7f800000 0x7ff00000 00000000 0x7fff000000000000

0000000000000000

-Inf

or -Infinity

0xff800000 0xfff00000 00000000 0xffff000000000000

0000000000000000

7.18 Magnetic Tape File I/OThe magnetic tape file I/O is explained here.

When the unit number is connected with the magnetic tape file by the environment variable or the OPEN statement, tape file I/O is possible.

The following I/O statement can be used.

- Unformatted sequential I/O statement

- Formatted sequential I/O statement

- File connection statement


- File inquiry statement

7.18.1 Connection of Magnetic Tape File and Unit NumberThe method of connecting the magnetic tape file and the unit number is explained.

7.18.1.1 Connection by Environment VariableThe magnetic tape file can be connected with the unit number by the environment variable.

Example: The magnetic tape file is connected with unit number 1.

Program:

OPEN(1) ...

Command string:

$ export fu01="/dev/rmt/0n"$ ./a.out

7.18.1.2 Connection by FILE Specifier of OPEN StatementA FILE specifier of the OPEN statement can connect magnetic tape file with the unit number.

Example: The magnetic tape file is connected with unit number 10.

Program:

OPEN(10,FILE='/dev/rmt/0n') ...

7.18.2 Note of Magnetic Tape File I/O FunctionHere, the note in the magnetic tape file I/O function use is explained.

- 167 -

- When the program terminates abnormally, the place where the tape is located might become uncertain.Please locate the tape at an appropriate position by using the mt(1) command.

- Executing option -g is ineffectual.

7.19 Conversion between EBCDIC Character Code and ASCIICharacter Code on Input/output

This section describes the conversion between IBM370 EBCDIC character code data and ASCII character code data.

7.19.1 Relationship between EBCDIC Character Code Data and Input/outputStatements

You may specify the runtime option -G to read and write IBM370 EBCDIC character code data. If the runtime option -G is specified, thefollowing operations are executed. See Section 3.3 Runtime Options for information on the runtime option -G.

The flow of IBM370 EBCDIC character code data.




Files written with variable-length Fortran records by Fujitsu Fortran implementation can be accessed by an unformatted input statement.

Character type of input/output list items are allowed.

7.20 Asynchronous Input/output StatementsThis section describes asynchronous input/output statements.

- 168 -

7.20.1 Execution of Asynchronous Input/output StatementsAn input/output statement that includes an ASYNCHRONOUS= specifier with a YES value is executed as an asynchronous input/outputstatement when it is used in relation to a file connected via an ASYNCHRONOUS= specifier with a YES value specified in an OPENStatement.

An asynchronous input/output statement is processed by making data transfer asynchronous for any data that can be transferredasynchronously within the system.

Note, however, that asynchronous data transfer is not enabled for input/output statements other than those for unformatted sequentialaccess or when the following compiler options or runtime options are specified:

- Compiler option -H ( see Section 2.2 Compiler Options)

- Compiler option -Eg (see Section 2.2 Compiler Options)

- Runtime option -G (see Section 3.3 Runtime Options)

- Runtime option -C (see Section 3.3 Runtime Options)

- Runtime option -T (see Section 3.3 Runtime Options)

In addition, asynchronous data transfer may not be enabled depending on the number, size, kind, type, or attributes of items that have beenspecified in an input/output list.

7.20.2 ID= Specifier in Data Transfer Input/output StatementsA literal value less than 9223372036854775808 more than 0 that uniquely identifies an asynchronous input/output statement executedduring the course of a program is assigned to an ID= specifier in a data transfer input/output statement. When a value that has been assignedto an ID= specifier exceeds the allowed range of ID= specifier values, that value is not guaranteed.

- 169 -

Chapter 8 DebuggingThis chapter describes the system debugging functions provided for finding and correcting errors in source programs.

8.1 Error ProcessingIf an error occurs during the execution of a Fortran program, the error monitor controls the following operations:

- Generation of a diagnostic message

- Continuation of execution if allowed by the error type

By using the error monitor, the user can control the following:

- The error count limit for canceling program execution

- The printed message count limit

- Whether to print the trace back map

- Whether to call user-defined error subroutines

If the user does not specify these control operations, the system uses its default operations. For details, see "Table 8.1 Error control tablestandard values" in the next section.

8.1.1 Error MonitorIf an error occurs during execution, the error monitor is called. The error monitor processes errors based on the error control table.

The error monitor does the following:

- Checks whether the error count limit has been reached and notifies the user if it has.

- Checks whether the printed message count limit has been reached and prints a diagnostic message if it has not.

- Generates a trace back map if requested. If the Fortran program is linked to a program written in another language, some trace backmap information may not be generated.

- Terminates the job if the error count limit has been reached. If the limit has not been reached, corrective action is taken and executioncontinues.

Corrective action for errors consists of standard corrections provided by the system and the execution of user-defined subroutines.

The figure below shows how the error monitor is called after an error.

- 170 -

Error monitor operations are based on data in the error control table. The error control table consists of 12-byte error items, as follows:

The error items are:

estop : At byte 0, an error count limit in the range 0 to 255. Zero indicates that unlimited errors are allowed.

mprint : At byte 1, a printed message limit in the range 0 to 255. Zero indicates that no messages are printed.

ecount : At byte 2, an error count during execution, in the range 0 to 255.

inf : At byte 3, an error control item, as follows:

Bit 0 : (1) The control character is set to the diagnostic messages. (0) The control character is not setto the diagnostic messages. See Section "8.1.2.3 ERRSET Subroutine".

Bit 1 : (1) The user can correct the error item. (0) The user cannot correct the error item.

Bit 2 : (1) The error count exceeds 255. (0) The error count does not exceed 255.

Bit 3 : (1) The buffer is printed. (0) The buffer is not printed.

Bit 4 : Not used

Bit 5 : (1) Messages are printed. (0) Message printing is based on the value of mprint.

Bit 6 : (1) The trace back map is printed. (0) The trace back map is not printed.

- 171 -

Bit 7 : Not used

uexit : At byte 4, the address of a user-defined error subroutine. If standard error correction is used, the last bit isset to 1.

ecode : At byte 8, an error level, as follows:

0 : i level. Fortran program processing continues after the message is printed.

4 : w level. Fortran program processing continues after the error is corrected.

8 : e level. The error statement is ignored. Corrective action is taken and execution continues ifthe error count limit has not been reached. The default error count limit is 10.

12 : s level. The error statement is ignored. Corrective action is taken and execution continues ifthe error count limit has not been reached. The default error count limit is 1.

16 : u level. Fortran program processing terminates.

Some error items cannot be edited. For editable items, use the ERRSET subroutine to change the standard values in the error control table.The Error Control Table Standard Values indicates whether error items are editable, and gives the default values for each item in the table.

Table 8.1 Error control table standard values

Error itemnumber

Error countlimit

Messagecount limit

Error itemeditable

Buffer Trace backmap

Standardcorrection

Errorlevel

11 to 14 1 1 Not Editable Not Printed Not Printed None u


19 1 1 Not Editable Not Printed Printed None u

20 1 1 Not Editable Not Printed Not Printed None u

21 to 22 1 1 Editable Not Printed Printed Yes s

28 1 1 Editable Not Printed Printed Yes s

42 10 5 Editable Not Printed Printed Yes e



64 to 65 10 5 Editable Not Printed Printed Yes e




83 to 86 Unlimited 5 Editable Not Printed Printed Yes w


97 Unlimited 5 Editable Not Printed Printed Yes w








131 to

13410 5

Editable Not Printed Printed Yes e


- 172 -

Error itemnumber

Error countlimit

Messagecount limit

Error itemeditable


Standardcorrection

Errorlevel


152 Unlimited 1 Editable Not Printed Not Printed Yes w



158 to 159 Unlimited 1 Editable Not Printed Not Printed Yes w



173 to 175 Unlimited 5 Editable Printed Printed Yes w


182 to 183 Unlimited 1 Editable Not Printed Not Printed Yes w

184 to 189 10 5 Editable Printed Printed Yes e






291 10 5 Editable Not Printed Printed None e


301 10 5 Editable Not Printed Printed Yes w




311 to 318 Unlimited Unlimited Editable Not Printed Printed Yes w





330 Unlimited Unlimited Editable Not Printed Printed Yes w





1033 Unlimited 5 Editable Not Printed Printed Yes i







- 173 -

Error itemnumber

Error countlimit

Messagecount limit

Error itemeditable


Standardcorrection

Errorlevel


1043 Unlimited 5 Editable Not Printed Not Printed Yes i

1044 to 1046 1 1 Not Editable Not Printed Printed None u



















1201

to 1203Unlimited 5

Editable Printed Printed Yes w

















- 174 -

Error itemnumber

Error countlimit

Messagecount limit

Error itemeditable


Standardcorrection

Errorlevel










1561 Unlimited Unlimited Editable Not Printed Printed Yes w









1652 Unlimited Unlimited Editable Not Printed Not Printed Yes w

If an error occurs, the error monitor prints the following message (indicating correction status) after the trace back map:

taken to (corrector) corrective action, execution continuing.

corrector: {standard|user}

standard: indicates that an error is corrected with standard correction.

user: indicates that an error is corrected with user-defined correction.

8.1.2 Error SubroutinesSeveral service subroutines provide user control of error processing during execution. These subroutines can be used to optimize errormonitor use.

With these subroutines, you can dynamically change items in the error control table, generate a trace back map, and execute user-definederror subroutines.

8.1.2.1 ERRSAV SubroutineERRSAV stores the leading 16 bytes of the specified error item in 16 bytes user area.

The ERRSAV service subroutine calling format is as follows:

CALL ERRSAV ( errno , darea )

errno

Default integer scalar. Error number.

darea

16 bytes character scalar. The error item is saved.

- 175 -

If the type of darea is not character, this result may be undefined.

8.1.2.2 ERRSTR SubroutineERRSTR moves an error item to the error control table element for the specified error number. Thus, when an error occurs, the new erroritem controls error processing.

The ERRSTR service subroutine calling format is as follows:

CALL ERRSTR ( errno , darea )

errno


darea

16 bytes character scalar. The error item is set.

If the type of darea is not character, this result may be undefined.

Example: ERRSTR and ERRSAV

CHARACTER(LEN=16) ERR113 !(1)ESTOP,(2)MPRINT,(3)ECOUNT,(4)INFCALL ERRSAV(113,ERR113)WRITE(ERR113(1:2),'(2a1)')0,0CALL ERRSTR(113,ERR113)OPEN(10,FILE='X.DAT',FORM='FORMATTED')DO I=1,15 WRITE(10)IEND DOCLOSE(10,STATUS='DELETE')END

8.1.2.3 ERRSET SubroutineERRSET changes the control table data for the error item with the specified error number. The error count limit, message print limit, traceback map printing switch, and error corrective action may be changed.

The ERRSET service subroutine calling format is as follows:

CALL ERRSET ( errno , estop , mprint , trace , uexit , r )

errno


estop

Default integer scalar. Error count limit.

<=0 : Error count limit is the same as before.

>=256 : It indicates unlimited error counts.

mprint

Default integer scalar. Message print limit.

<0 : Message print limit is zero (that is, a no-print message).

=0 : Message print limit is the same as before.

>=256 : It indicates unlimited message prints.

trace

Default integer scalar. One of the following values indicates whether to print a trace back map:

=0 : Do not change.

=1 : Do not print.

- 176 -

=2 : Print.

uexit

Default integer scalar. Error correction or user-defined correction subroutine indicated by one of the following:

=0 : Do not change.

=1 : Perform standard correction.

=other : Execute the user-defined correction.

r

If errno is not 132, a four-byte integer expression indicates the largest error number to be changed.

If errno is 132, indicates whether a blank character is inserted at the beginning of the correction data with the following values:

=1 : Insert a blank character.

=other : Do not insert a blank character.

Example: ERRSET Service Subroutine

EXTERNAL FIXUPCALL ERRSET(202,50,-1,0,FIXUP,205)

In this example, the error items for error numbers 202 to 205 are changed as follows:

- The error count limit is changed to 50.

- A message will not be printed if an error occurs.

- The print control information for the trace back map is not changed.

- The user-defined subroutine FIXUP is called to process errors.

8.1.2.4 ERRTRA SubroutineThe ERRTRA service subroutine generates a trace back map up to the program unit currently executing. See Section "4.2 ExecutableProgram Output" for the map format. Execution continues after this subroutine is called.

The ERRTRA service subroutine calling format is as follows:

CALL ERRTRA

Example: ERRTRA Service Subroutine

OPEN(10,FILE='X.DAT')CALL SUB1ENDSUBROUTINE SUB1CALL ERRTRADO I=1,15 WRITE(10,*)IEND DOEND

8.1.3 Using the Error MonitorBy calling the error monitor, the user can control error processing, specify user-defined error subroutines, and use the error service functions.Related topics explain how to use the error service functions.

8.1.3.1 User-Defined Error SubroutinesThe ERRSET subroutine may be used to store the address of a user-defined error subroutine in the error control table. If a user-definederror subroutine has been specified, the error monitor calls the error subroutine as follows:

- 177 -

CALL user ( ret , errno [ , data1 , ... , datan ] )

user : Name of the user-defined error subroutine.

ret : four-byte integer scalar variable that stores a return code. The return code must be provided by the user-defined error subroutine.

errno : Number of the detected error.

data1, ..., datan : Scalar variable names or array element name list used by the user-defined subroutine.

The arguments passed to the user-defined error subroutine are as follows:

- Input-output error: The arguments in Section "8.1.4 Input/Output Error Processing" are passed.

- Intrinsic function error: The arguments in Section "8.1.5 Intrinsic Function Error Processing" are passed.

- Intrinsic subroutine error: The arguments in Section "8.1.6 Intrinsic Subroutine Error Processing" are passed.

- Interrupt error: The arguments in Section "8.1.7 Exception Handling Processing" are passed.

- Otherwise, the arguments in Section "8.1.8 Other Error Processing" are passed.

The following restrictions apply to user-defined error subroutines:

- Restrictions on input/output statements.If an error occurs, input/output statements cannot be used with the same unit number as the unit with the error. For example, if anerror occurs with unit number 10, input/output statements with unit number 10 cannot be used.

- Restrictions on intrinsic functions.If an error occurs, the intrinsic function that caused the error cannot be used in the user-defined subroutine. For example, if an argumentvalue is incorrect in SQRT, SQRT is disabled in the user-defined subroutine. This restriction also applies to intrinsic functionsreferenced implicitly when expressions that include exponentiation are used.

- Restrictions on intrinsic subroutines.If an error occurs, the intrinsic subroutine that caused the error cannot be used in the user-defined subroutine.

- Restrictions on return codes.If arguments data1...datan of the user-defined error subroutine are not changed, the user-defined subroutine must set the return codeto 0. If the passed arguments are changed, the subroutine must set the return code to 1. Only 0 or 1 may be used as return code values.For all other values, the results are unpredictable.

- Restrictions on the argument type.For error subroutines written in Fortran, the argument type must match that of the CALL statement.

8.1.3.2 Handling of Incorrect Arguments in Intrinsic FunctionsIf an intrinsic function argument is incorrect, the argument is assigned to a local variable in the intrinsic function and the error monitor iscalled. The local variable is passed to the error monitor. In user-defined error subroutines, this local variable must be corrected. Do notcorrect the argument in the program unit that used the incorrect argument.

8.1.3.3 User Error ProcessingThe following example shows user error processing. In this example, when the user enters incorrect data (1234567890123), error number171 is generated. To continue processing, the user corrects the data to 9999 with the user-defined error subroutine FIXUP.

Example: User error processing

Fortran program:

EXTERNAL FIXUP CALL ERRSET(171,0,0,0,FIXUP,0) READ(5,*)I WRITE(6,100)I 100 FORMAT(1H, 5X, 'I=',I4) STOP END

User-defined error correction subroutine:

- 178 -

SUBROUTINE FIXUP(IRET,INO,I4CONV)IF(INO==171)THEN I4CONV=9999 IRET=1END IFRETURNEND

Input data:

1234567890123

Result:

jwe0171i-e line 3 Integer(kind=1), integer(kind=2) or integer(kind=4) number (1234567890123)is out of range (unit= 5). error occurs at MAIN__ line 3 loc 0000000000400bb5 offset 0000000000000055 MAIN__ at loc 0000000000400b60 called from o.s.taken to (user) corrective action, execution continuing. I=9999

8.1.4 Input/Output Error ProcessingIf an IOSTAT= specifier, ERR= specifier, END= specifier, or EOR= specifier has been specified, input/output errors are processed basedon the specifier without regard to the error count limit. See Section "7.7.1.3 IOSTAT= Specifier", Section "7.7.1.4 Error Branch", Section"7.7.1.5 End-of-File Branch", and Section "7.7.1.6 End-of-Record Branch" for details.

If an error occurs, the file position and input list values being executed at the time of the error are undefined.

The table below lists the standard system and user-defined corrections for errors that occur during input/output statement execution. Thefollowing arguments are passed to the user-defined error program:

- Return code

- Error number

- The data listed in the table below

Table 8.2 Error processing for input/output

Errornumber

Standard correction processing User-defined errorsubroutine

Data passed to user-defined subroutine

Data1 Data2

21 to 22 Terminates the program Not correctable UNIT None

28 Terminates the program Not correctable UNIT None

42 Ignores the statement Not correctable UNIT None



64 to 65 Ignores the statement Not correctable UNIT None



80 Ignores the statement Not correctable UNIT REC

81 Ignores the statement Not correctable None None


83 to 85 Continues processing without setting a specifier value. Not correctable None None

86 Continues processing without setting a specifier value. Not correctable UNIT None

- 179 -

Errornumber



Data1 Data2


97 Uses DIRECT as the ACCESS=specifier Not correctable UNIT None

102 to103

Ignores the statement Not correctable UNIT None

104 Ignores a POSITION=specifier and continuesprocessing

Not correctable UNIT None


109 The maximum value of the type specified by specifiervalue is set, and processing is continued.


111 to115




131 Ignores subsequent input/output lists Not correctable None None

132 a. For input, ignores data that exceeded the record

b. For output, writes the excess data to the nextlogical record


133 Ignores subsequent output list items Not correctable UNIT None

134 Ignores subsequent input/output lists Not correctable UNIT None

140 Ignores the conversion and continues processing Not correctable UNIT None

141 to142

Assigns maximum value after transforming data Assign thetransformed data toFDATA

FDATA UDATA

143 to144

Assigns the value 0 after transforming data Assign thetransformed data toFDATA

FDATA UDATA

145 to146

Assigns maximum value (Inf or NaN) aftertransforming data

Assign thetransformed data toFDATA

FDATA UDATA

147 The last bit of mantissa assigns the value Not correctable UNIT None

151 Ignores subsequent input/output lists Not correctable UNIT None

152 Edits according to the edit descriptor Not correctable UNIT None

153 Adds the initial left parentheses Not correctable UNIT None

154 Ignores nests exceeding 30 levels in the formatspecification


155 Ignores the edit descriptor and goes by default to thefinal right parenthesis of the format specification


156 Ignores the repeat specification and goes by default tothe final right parenthesis of the format specification


157 Ignores the incorrect character and goes by default tothe final right parenthesis of the format specification

Assign the characterof the formatspecification toFCODE

FCODE None

- 180 -

Errornumber



Data1 Data2

158 Adds the right parenthesis and continues formatcontrol


159 Replaces the input data with the character string in theformat specification


162 Ignores the input/output list and continues a formatcontrol


171 a. Uses by default the minimum specifiable valueif the one-, two- or four-byte input item isnegative

b. Uses by default the maximum specifiable valueif the one-, two- or four-byte input item ispositive

Correct I4CNV I4CNV None

172 a. Uses by default the minimum specifiable valueif the eight-byte input item is negative

b. Uses by default the maximum specifiable valueif the eight-byte input item is positive

Correct I8CNV I8CNV None

173 to175

Uses 0 for the incorrect character Assign the correctedcharacter to IDATA

IDATA None

176 a. Sets true 0 if the absolute value is less than2**(-126)*2**(-23) (real),2**(-1022)*2**(-52) (double-precision real) or2**(-16382)*2**(-112) (quadruple-precisionreal)

b. Sets 2**127*(1+1.0-2**(-23))(real),2**1023*(1+1.0-2**(-52)) (double-precisionreal) or 2**16383*(1+1.0- 2**(-112))(quadruple-precision real) without changing thesign if the absolute value is greater than2**127*(1+1.0-2**(-23)) real),2**1023*(1+1.0-2**(-52)) (double-precisionreal) or 2**16383*(1+1.0-2**(-112))(quadruple-precision real)

Correct FCONV FCONV None

182 to183

Executes character editing Not correctable UNIT None

184 Ignores the constant and terminates the statement Not correctable UNIT None

185 Ignores the repeat count or constant and terminates thestatement


186 to188

Ignores the item name and terminates the statement Not correctable UNIT None


190 Ignores the constant that exceeded the element andcontinues the processing



1111 to1113


- 181 -

Errornumber



Data1 Data2

1114 Assigns the maximum value for 4 byte integer to thespecifier, and continues the processing


1115 An invalid specifier in an open statement, andcontinues the processing


1117 to1122





1149 Uses NO as the ASYNCHRONOUS= specifier Not correctable UNIT None


1161 Terminates the program Not correctable UNIT None

1162 The multiple records are written or read, and continuesthe processing


1163 Assigns the maximum value for 4 byte integer to thevariable, and continues the processing

Not correctable None None

1181 Ignores the input data and continues the program Not correctable UNIT None

1201 Uses 0 for the incorrect character Assign the correctedcharacter to IDATA

IDATA None

1202 Ignores the input data and continues the processing Not correctable UNIT None

1203 Edits with the type of data and continues theprocessing


UNIT: For an external file, the unit number (four-byte integer) is specified. For an internal file, -1 is specified.

IDATA: Incorrect character in the input data (one-byte character)

FCODE: Incorrect format code (one-byte character) in the format assigned by the array

REC: Record variable (four-byte integer)

I4CNV: Integer indicating result of format conversion (four-byte integer)

I8CNV: Integer indicating result of format conversion (eight-byte integer)

FCONV: Integer indicating result of format conversion (real)

FDATA: Integer indicating result of IEEE-IBM370 conversion (real)

UDATA: For an input/output statement, the same as UNIT; for a service subroutine, the maximum value of a four-byte integer

8.1.5 Intrinsic Function Error ProcessingThe table below lists the causes of intrinsic function errors and the standard system correction processing in execution.

When -Ncheck_intrfunc option is specified, the error numbers of 202-255, 259-282, 1355-1370, 1396-1414, 1416-1417, and 1425-1454are detected and the intrinsic function error processing is performed.

When -Ncheck_intrfunc option is not specified, the floating point exception occurs instead of the error. The floating point exception isdetected that -NRtrap option and environment variable FLIB_EXCEPT=u are specified.

See Section "2.2 Compiler Options" for details of -Ncheck_intrfunc option. See Section "3.8 Using Environment Variables forExecution" for detail of environment variable FLIB_EXCEPT=u.

- 182 -

The following arguments are passed to the user-defined error subroutine:

- Return code

- Error number


Table 8.3 Error processing for intrinsic function

Errornumber

Reference format Cause of error Standard correctionprocessing

Data passed touser-definedsubroutine

data1 data2

202 SIN(X) |X|>=8.23e+05 NaN X None

COS(X)

203 DSIN(DX) |DX|>=3.53d+15 NaN DX None

DCOS(DX)

204 QSIN(QX) |QX|>=2**62*PI NaN QX None

QCOS(QX)

205 CSIN(CX) |X1|>=8.23e+05 (NaN,NaN) X1 None

CCOS(CX)

206 CSIN(CX) X2>=89.415 (SIGN(Inf,SIN(X1)),SIGN(Inf,COS(X1)))

X2 None

X2<=-89.415 (SIGN(Inf,SIN(X1)),-SIGN(Inf,COS(X1)))

CCOS(CX) X2>=89.415 (SIGN(Inf,COS(X1)),-SIGN(Inf,SIN(X1)))

X2<=-89.415 (SIGN(Inf,COS(X1)),SIGN(Inf,SIN(X1)))

207 CDSIN(CDX) |DX1|>=3.53d+15 (NaN,NaN) DX1 None

CDCOS(CDX)

208 CDSIN(CDX) DX2>=710.475 (DSIGN(Inf,DSIN(DX1)),DSIGN(Inf,DCOS(DX1)))

DX2 None

DX2<=-710.475 (DSIGN(Inf,DSIN(DX1)),-DSIGN(Inf,DCOS(DX1)))

CDCOS(CDX) DX2>=710.475 (DSIGN(Inf,DCOS(DX1)),-DSIGN(Inf,DSIN(DX1)))

DX2<=-710.475 (DSIGN(Inf,DCOS(DX1)),DSIGN(Inf,DSIN(DX1)))

209 CQSIN(CQX) |QX1|>=2**62*PI (NaN,NaN) QX1 None

CQCOS(CQX)

210 CQSIN(CQX) QX2>=11357.125 (QSIGN(Inf,QSIN(QX1)),QSIGN(Inf,QCOS(QX1)))

QX2 None

QX2<=-11357.125 (QSIGN(Inf,QSIN(QX1)),-QSIGN(Inf,QCOS(QX1)))

CQCOS(CQX) QX2>=11357.125 (QSIGN(Inf,QCOS(QX1)),-QSIGN(Inf,QSIN(QX1)))

- 183 -

Errornumber



data1 data2

QX2<=-11357.125 (QSIGN(Inf,QCOS(QX1)),QSIGN(Inf,QSIN(QX1)))

211 TAN(X) |X|>=8.23e+05 NaN X None

COTAN(X)

212 TAN(X) |X| is close tosingularity (+-PI/2,+-3PI/2,...)

Inf X None

COTAN(X) |X| is close tosingularity (0,+-PI,+-2PI,...)

213 DTAN(DX) |DX|>=3.53d+15 NaN DX None

DCOTAN(DX)

214 DTAN(DX) |DX| is close tosingularity (+-PI/2,+-3PI/2,...)

Inf DX None

DCOTAN(DX) |DX| is close tosingularity (0,+-PI,+-2PI,...)

215 QTAN(QX) |QX|>=2**62*PI NaN QX None

QCOTAN(QX)

216 QTAN(QX) |QX| is close tosingularity (+-PI/2,+-3PI/2,...)

Inf QX None

QCOTAN(QX) |QX| is close tosingularity (0,+-PI,+-2PI,...)

Inf QX None

217 ASIN(X) |X|>1.0 NaN X None

ACOS(X)

218 DASIN(DX) |DX|>1.0 NaN DX None

DACOS(DX)

219 QASIN(QX) |QX|>1.0 NaN QX None

QACOS(QX)

220 ATAN(X1,X2) X1=0.0 andX2=0.0

NaN X1 X2

ATAN2(X1,X2)

221 DATAN2(DX1,DX2) DX1=0.0 andDX2=0.0

NaN DX1 DX2

222 QATAN2(QX1,QX2) QX1=0.0 andQX2=0.0

NaN QX1 QX2

223 SINH(X) X>=89.415 Inf X None

X<=-89.415 -Inf

COSH(X) |X|>=89.415 Inf

- 184 -

Errornumber



data1 data2

224 DSINH(DX) DX>=710.475 Inf DX None

DX<=-710.475 -Inf

DCOSH(DX) |DX|>=710.475 Inf

225 QSINH(QX) QX>=11357.125 Inf QX None

QX<=-11357.125 -Inf

QCOSH(QX) |QX|>=11357.125 Inf

226 SQRT(X) X<0.0 NaN X None

227 DSQRT(DX) DX<0.0 NaN DX None

228 QSQRT(QX) QX<0.0 NaN QX None

229 EXP(X) X>=88.722 Inf X None

230 DEXP(DX) DX>=709.782 Inf DX None

231 QEXP(QX) QX>=11356.5 Inf QX None

232 CEXP(CX) X1>=88.722 (SIGN(Inf,COS(X1)),SIGN(Inf,SIN(X1)))

X1 None

233 CEXP(CX) |X2|>=8.23e+05 (NaN,NaN) X2 None

234 CDEXP(CDX) DX1>=709.782 (DSIGN(Inf,DCOS(DX1)),DSIGN(Inf,DSIN(DX1)))

DX1 None

235 CDEXP(CDX) |DX2|>=3.53d+15 (NaN,NaN) DX2 None

236 CQEXP(CQX) QX1>=11356.5 (QSIGN(Inf,QCOS(QX1)),QSIGN(Inf,QSIN(QX1)))

QX1 None

237 CQEXP(CQX) |QX2|>=2**62*PI (NaN,NaN) QX2 None

238 EXP2(X) X>=128.0 Inf X None

2.0**X

239 DEXP2(DX) DX>=1024.0 Inf DX None

2.0**DX

240 QEXP2(QX) QX>=16384.0 Inf QX None

2.0**QX

241 EXP10(X) X>=38.531 Inf X None

10.0**X

242 DEXP10(DX) DX>=308.254 Inf DX None

10.0**DX

243 QEXP10(QX) QX>=4932.0625 Inf QX None

10.0**QX

244 ALOG(X) X=0.0 -Inf X None

X<0.0 NaN

ALOG10(X) X=0.0 -Inf

X<0.0 NaN

ALOG2(X) X=0.0 -Inf

- 185 -

Errornumber



data1 data2

X<0.0 NaN

245 DLOG(DX) DX=0.0 -Inf DX None

DX<0.0 NaN

DLOG10(DX) DX=0.0 -Inf

DX<0.0 NaN

DLOG2(DX) DX=0.0 -Inf

DX<0.0 NaN

246 QLOG(QX) QX=0.0 -Inf QX None

QX<0.0 NaN

QLOG10(QX) QX=0.0 -Inf

QX<0.0 NaN

QLOG2(QX) QX=0.0 -Inf

QX<0.0 NaN

247 CLOG(CX) CX=(0.0,0.0) (Inf,NaN) CX None

248 CDLOG(CDX) CDX=(0.0,0.0) (Inf,NaN) CDX None

249 CQLOG(CQX) CQX=(0.0,0.0) (Inf,NaN) CQX None

250 GAMMA(X) X<=0.0 NaN X None

X>=35.039860 Inf

251 DGAMMA(DX) DX<=0.0 NaN DX None

DX>=171.6243 Inf

252 QGAMMA(QX) QX<=0.0 NaN QX None

QX>=1.755q+03 Inf

253 ALGAMA(X) X<=0.0 NaN X None

X>=0.403711e+37 Inf

254 DLGAMA(DX) DX<=0.0 NaN DX None

DX>=2.55634d+305

Inf

255 QLGAMA(QX) QX<=0.0 NaN QX None

QX>=1.048q+4928

Inf

256 LGT(C1,C2) An EBCDIC code,not correspondingto an ASCII codewas detected in acharacter string C1or C2

The default ASCII collationorder of incorrect argument,26(X'1A'), is used.

C3 None

LGE(C1,C2)

LLT(C1,C2)

LLE(C1,C2)

257 IBSET(I,POS) The value of POSexceeded the range

I None

IBCLR(I,POS)

BTEST(I,POS) .FALSE.

- 186 -

Errornumber



data1 data2

259 IX1**IX2 IX1=0 and IX2<0 0 IX1 IX2

260 JX1**JX2 JX1=0 and JX2<0 0 JX1 JX2

261 X**IX X=0.0 and IX<0 Inf X IX

262 X1**X2 X1=0.0 andX2<0.0

Inf X1 X2

263 X1**X2 X1<0.0 and X2/=0.0

NaN X1 X2

264 DX**IX DX=0.0 and IX<0 Inf DX IX

265 DX**JX DX=0.0 and JX<0 Inf DX JX

266 DX1**DX2 DX1=0.0 andDX2<0.0

Inf DX1 DX2

267 DX1**DX2 DX1<0.0 andDX2/=0.0

NaN DX1 DX2

268 QX**IX QX=0.0 and IX<0 Inf QX IX

269 QX**JX QX=0.0 and JX<0 Inf QX JX

270 QX1**QX2 QX1=0.0 andQX2<0.0

Inf QX1 QX2

271 QX1**QX2 QX1<0.0 andQX2/=0.0

NaN QX1 QX2

272 QX1**QX2 |QX1**QX2|>=2**16384

Inf QX1 QX2

273 CX**IX CX=(0.0,0.0) andIX<0

(NaN,NaN) CX IX

274 CDX**IX CDX=(0.0,0.0) andIX<0

(NaN,NaN) CDX IX

275 CDX**JX CDX=(0.0,0.0) andJX<0

(NaN,NaN) CDX JX

276 CQX**IX CQX=(0.0,0.0) andIX<0

(NaN,NaN) CQX IX

277 CQX**JX CQX=(0.0,0.0) andJX<0

(NaN,NaN) CQX JX

278 QX1/QX2 |QX1/QX2|>=2**16384

Inf QX1 QX2

279 QX1/QX2 |QX1/QX2|<2**(-16382)

0.0 QX1 QX2

280 QX1/QX2 QX2=0.0 For QX1=0.0NaN

For QX1/=0.0SIGN(QX1)*Inf

QX1 QX2

281 JX1+JX2 |JX1+JX2|>2**63-1

2**63-1 JX1 JX2

- 187 -

Errornumber



data1 data2

COTAND(X) |X| is close tosingularity (0,+-180,+-360, ...)

1366 DTAND(DX) |DX| is close tosingularity (+-90,+-270, ...)

Inf DX None

DCOTAND(DX) |DX| is close tosingularity (0,+-180,+-360, ...)

1367 ASIND(X) |X|>1.0 NaN X None

ACOSD(X)

1368 DASIND(DX) |DX|>1.0 NaN DX None

DACOSD(DX)

1369 ATAN2D(X1,X2) X1=0.0 andX2=0.0

NaN X1 X2

1370 DATAN2D(DX1,DX2) DX1=0.0 andDX2=0.0

NaN DX1 DX2

1371 ISHFTC(I,SHIFT[,SIZE]) SIZE<=0 orSIZE>BIT_SIZE(I)

BIT_SIZE(I) is set to SIZE,and process is continued

None

|SHIFT|>SIZE Value of SIZE is set toSHIFT, and process iscontinued

1372 IBITS(I,POS,LEN) LEN<0 orLEN>BIT_SIZE(I)-POS

0 None

POS<0 orPOS>=BIT_SIZE(I)-LEN

1373 ISHFT(I,SHIFT) |SHIFT|>BIT_SIZE(I)

BIT_SIZE(I) is set to SHIFT,and process is continued

None

1375 MATMUL(MATRIX_A,MATRIX_B)

The size of lastdimension ofMATRIX_A andthe size of firstdimension ofMATRIX_B is notsame

Terminates the program None

1376 SPREAD(SOURCE,DIM,NCOPIES)

DIM<=0 orDIM>(The rank ofSOURCE)+1


1377 ALL(MASK[,DIM]) DIM<=0 orDIM>(The rank ofMASK)


ANY(MASK[,DIM])

COUNT(MASK[,DIM,KIND])

- 189 -

Errornumber



data1 data2

SUM(ARRAY,DIM[,MASK]) DIM<=0 orDIM>(The rank ofARRAY)

PRODUCT(ARRAY,DIM[,MASK])

MAXVAL(ARRAY,DIM[,MASK])

MINVAL(ARRAY,DIM[,MASK])

UBOUND(ARRAY[,DIM])

LBOUND(ARRAY[,DIM,KIND])

SIZE(ARRAY[,DIM])

MAXLOC(ARRAY,DIM[,MASK,KIND])

MINLOC(ARRAY,DIM[,MASK,KIND])

1378 MAXLOC(ARRAY,DIM[,MASK,KIND])

or

MAXLOC(ARRAY[,MASK,KIND])

The shape ofARRAY andMASK is different


MINLOC(ARRAY,DIM[,MASK,KIND])

or

MINLOC(ARRAY[,MASK,KIND])

SUM(ARRAY,DIM[,MASK])

or

SUM(ARRAY[,MASK])

PRODUCT(ARRAY,DIM[,MASK])

or

PRODUCT(ARRAY[,MASK])

MAXVAL(ARRAY,DIM[,MASK])

or

MAXVAL(ARRAY[,MASK])

MINVAL(ARRAY,DIM[,MASK])

or

MINVAL(ARRAY[,MASK])

PACK(ARRAY,MASK[,VECTOR])

- 190 -

Errornumber



data1 data2

1381 PACK(ARRAY,MASK[,VECTOR])

The elementsnumber ofVECTOR is lessthan total numberof TRUE of MASK


1382 UNPACK(VECTOR,MASK[,FIELD])

The elementsnumber ofVECTOR is lessthan total numberof TRUE of MASK


1383 UNPACK(VECTOR,MASK[,FIELD])

The shape ofMASK and FIELDis not same


1384 CSHIFT(ARRAY,SHIFT[,DIM]) DIM<=0 orDIM>(The rank ofARRAY)


EOSHIFT(ARRAY,SHIFT[,BOUNDARY][,DIM])

1385 CSHIFT(ARRAY,SHIFT[,DIM]) The shape ofSHIFT is invalid



1386 EOSHIFT(ARRAY,SHIFT[,BOUNDARY][,DIM])

The shape ofBOUNDARY isinvalid


1387 RESHAPE(SOURCE,SHAPE[,PAD][,ORDER])

The size ofSOURCE is lessthan the product ofevery element'svalues of SHAPE



The shape ofSHAPE andORDER isdifferent



The value of theelements ORDERmust be thecombination of thenumbers to the sizeof SHAPE


1392 REPEAT(STRING,NCOPIES) NCOPIES<0 Terminates the program None

1393 NEAREST(X,S) S=0.0 X None

1394 SIZE(ARRAY[,DIM]) If ARRAY is theassumed-size arrayDIM<1 orDIM>=(The rankof ARRAY)


UBOUND(ARRAY[,DIM])


The type parameterof SOURCE andPAD is different


- 191 -

Errornumber



data1 data2

PACK(ARRAY,MASK[,VECTOR])

The type parameterof ARRAY andVECTOR isdifferent

UNPACK(VECTOR,MASK[,FIELD])

The type parameterof VECTOR andFIELD is different

MERGE(TSOURCE,FSOURCE,MASK)

The type parameterof TSOURCE andFSOURCE isdifferent


The type parameterof ARRAY andBOUNDARY isdifferent

1396 MODULO(A,P) P=0 For REAL type AInf

For INTEGER type AHUGE(A)

None

1397 QX1+QX2 |QX1+QX2|>=2**16384(overflow)

Inf QX1 QX2

QX1-QX2 |QX1-QX2|>=2**16384(overflow)

QX1*QX2 |QX1*QX2|>=2**16384(overflow)

1398 QX1*QX2 |QX1*QX2|<2**(-16382)(under flow)

0.0 QX1 QX2

1399 QTANQ(QX) |QX| is close tosingularity (+-1,+-3, ...)

Inf QX None

QCOTANQ(QX) |QX| is close tosingularity (0,+-2,+-4, ...)

1400 QASINQ(QX) |QX|>1.0 NaN QX None

QACOSQ(QX)

1401 QATAN2Q(QX1,QX2) QX1=0.0 andQX2=0.0

NaN QX1 QX2

1402 QSIND(QX) |QX|>=2**62*180 NaN QX None

QCOSD(QX)

1403 QTAND(QX) |QX|>=2**63*90 NaN QX None

QCOTAND(QX)

- 192 -

Errornumber



data1 data2

1404 QTAND(QX) |QX| is close tosingularity (+-90,+-270, ...)

Inf QX None

QCOTAND(QX) |QX| is close tosingularity (0,+-180,+-360, ...)

1405 QASIND(QX) |QX|>1.0 NaN QX None

QACOSD(QX)

1406 QATAN2D(QX1,QX2) QX1=0.0 andQX2=0.0

NaN QX1 QX2

1407 CSINQ(CX) |X1|>=5.24e+05 (NaN,NaN) X1 None

CCOSQ(CX)

1408 CSINQ(CX) X2>=56.92 (SIGN(Inf,SINQ(X1)),SIGN(Inf,COSQ(X1)))

X2 None

X2<=-56.92 (SIGN(Inf,SINQ(X1)),-SIGN(Inf,COSQ(X1)))

CCOSQ(CX) X2>=56.92 (SIGN(Inf,COSQ(X1)),-SIGN(Inf,SINQ(X1)))

X2<=-56.92 (SIGN(Inf,COSQ(X1)),SIGN(Inf,SINQ(X1)))

1409 CDSINQ(CDX) |DX1|>=2.25d+15 (NaN,NaN) DX1 None

CDCOSQ(CDX)

1410 CDSINQ(CDX) DX2>=452.30 (DSIGN(Inf,DSINQ(DX1)),DSIGN(Inf,DCOSQ(DX1)))

DX2 None

DX2<=-452.30 (DSIGN(Inf,DSINQ(DX1)),-DSIGN(Inf,DCOSQ(DX1)))

CDCOSQ(CDX) DX2>=452.30 (DSIGN(Inf,DCOSQ(DX1)),-DSIGN(Inf,DSINQ(DX1)))

DX2<=-452.30 (DSIGN(Inf,DCOSQ(DX1)),DSIGN(Inf,DSINQ(DX1)))

1411 CQSINQ(CQX) |QX1|>=2**63 (NaN,NaN) QX1 None

CQCOSQ(CQX)

1412 CQSINQ(CQX) QX2>=7230.125 (QSIGN(Inf,QSINQ(QX1)),QSIGN(Inf,QCOSQ(QX1)))

QX2 None

QX2<=-7230.125 (QSIGN(Inf,QSINQ(QX1)),-QSIGN(Inf,QCOSQ(QX1)))

CQCOSQ(CQX) QX2>=7230.125 (QSIGN(Inf,QCOSQ(QX1)),-QSIGN(Inf,QSINQ(QX1)))

- 193 -

Errornumber



data1 data2

QX2<=-7230.125 (QSIGN(Inf,QCOSQ(QX1)),QSIGN(Inf,QSINQ(QX1)))

1413 X**JX X=0.0 and JX<0 Inf X JX

1414 CX**JX CX=(0.0,0.0) andJX<0

(NaN,NaN) CX JX

1415 SELECTED_REAL_KIND(P,R) P and R are notspecified


1416 X1**X2 |X1**X2|>=2**128

Inf X1 X2

1417 DX1**DX2 |DX1**DX2|>=2**1024

Inf DX1 DX2

1418 PACK(ARRAY,MASK[,VECTOR])

The elementsnumber ofVECTOR must begreater than orequal to the totalnumber of ARRAYif MASK is scalarwith the valueTRUE


1419 EOSHIFT(ARRAY,SHIFT[,BOUNDARY][,DIM])

Do not omitBOUNDARY,when ARRAY is aderived type.


1420 ICHAR(C [,KIND]) The function resultexceeds theexpressible rangethat kind typeparameter isspecified by KIND

The kind type parameter ofresult uses the specified kindvalue

None

INDEX(STRING,SUBSTRING[,BACK, KIND])

LBOUND(ARRAY,[, DIM,KIND])

LEN(STRING[, KIND])

SCAN(STRING, SET[, BACK,KIND])

SHAPE(SOURCE[, KIND])

SIZE(ARRAY[,DIM, KIND])

UBOUND(ARRAY[, DIM,KIND])

VERIFY(STRING, STE[, BACK,KIND])

1421 IEEE_CLASS(X) The function usedis not supported.

0 or .FALSE. None

IEEE_COPY_SIGN(X,Y)

IEEE_IS_FINITE(X)

IEEE_IS_NORMAL(X)

IEEE_IS_NEGATIVE(X)

- 194 -

Errornumber



data1 data2

IEEE_LOGB(X)

IEEE_NEXT_AFTER(X,Y)

IEEE_REM(X,Y)

IEEE_RINT(X)

IEEE_SCALB(X, I)

IEEE_UNORDERD(X,Y)

IEEE_VALUE(X,CLASS)

1422 IEEE_IS_NAN(X) The function usedis not supported.

0 or .FALSE. None

IEEE_VALUE(X,CLASS)

1425 ACOSH(X) X<1 NaN X None

1426 ACOSH(DX) DX<1 NaN DX None

1427 ACOSH(QX) QX<1 NaN QX None

1428 ATANH(X) |X|>1 NaN X None

X=1 Inf

X=-1 -Inf

1429 ATANH(DX) |DX|>1 NaN DX None

DX=1 Inf

DX=-1 -Inf

1430 ATANH(QX) |QX|>1 NaN QX None

QX=1 Inf

QX=-1 -Inf

1431 TAN(CX) |X1|>=8.23E+05 (NaN,NaN) X1 None

1432 TAN(CX) CX is close tosingularity ((+-PI/2,0),(+-3PI/2,0), ...)

For REAL(CX)>=0.0(Inf,0.0)

For REAL(CX)<0.0(-Inf,0.0)

CX None

1433 TAN(CDX) |DX1|>=3.53D+15 (NaN,NaN) DX1 None

1434 TAN(CDX) CDX is close tosingularity ((+-PI/2,0),(+-3PI/2,0), ...)

For DREAL(CDX)>=0.0(Inf,0.0)

For DREAL(CDX)<0.0(-Inf,0.0)

CDX None

1435 TAN(CQX) |QX1|>=2.0**62*PI

(NaN,NaN) QX1 None

1436 TAN(CQX) CQX is close tosingularity ((+-PI/2,0),(+-3PI/2,0), ...)

For QREAL(CQX)>=0.0(Inf,0.0)

For QREAL(CQX)<0.0(-Inf,0.0)

CQX None

1437 ATAN(CX) CX=(0.0,+-1.0) (0.0,+-Inf) CX None

1438 ATAN(CDX) CDX=(0.0,+-1.0) (0.0,+-Inf) CDX None

- 195 -

Errornumber



data1 data2

1439 ATAN(CQX) CQX=(0.0,+-1.0) (0.0,+-Inf) CQX None

1440 SINH(CX) |X2|>=8.23E+05 (NaN,NaN) X2 None

COSH(CX)

1441 SINH(CX) X1>=89.4150 (SIGN(Inf,COS(X2)),SIGN(Inf,SIN(X2)))

X1 None

X1<=-89.4150 (-SIGN(Inf,COS(X2)),SIGN(Inf,SIN(X2)))

COSH(CX) X1>=89.4150 (SIGN(Inf,COS(X2)),SIGN(Inf,SIN(X2)))

X1<=-89.4150 (SIGN(Inf,COS(X2)),-SIGN(Inf,SIN(X2)))

1442 SINH(CDX) |DX2|>=3.53D+15 (NaN,NaN) DX2 None

COSH(CDX)

1443 SINH(CDX) DX1>=710.475 (DSIGN(Inf,DCOS(DX2)),DSIGN(Inf,DSIN(DX2)))

DX1 None

DX1<=-710.475 (-DSIGN(Inf,DCOS(DX2)),DSIGN(Inf,DSIN(DX2)))

COSH(CDX) DX1>=710.475 (DSIGN(Inf,DCOS(DX2)),DSIGN(Inf,DSIN(DX2)))

DX1<=-710.475 (DSIGN(Inf,DCOS(DX2)),-DSIGN(Inf,DSIN(DX2)))

1444 SINH(CQX) |QX2|>=2.0**62*PI

(NaN,NaN) QX2 None

COSH(CQX)

1445 SINH(CQX) QX1>=11357.125 (QSIGN(Inf,QCOS(QX2)),QSIGN(Inf,QSIN(QX2)))

QX1 None

QX1<=-11357.125 (-QSIGN(Inf,QCOS(QX2)),QSIGN(Inf,QSIN(QX2)))

COSH(CQX) QX1>=11357.125 (QSIGN(Inf,QCOS(QX2)),QSIGN(Inf,QSIN(QX2)))

QX1<=-11357.125 (QSIGN(Inf,QCOS(QX2)),-QSIGN(Inf,QSIN(QX2)))

1446 TANH(CX) |X2|>=8.23E+05 (NaN,NaN) X2 None

1447 TANH(CX) CX is close tosingularity ((0,+-PI/2),(0,+-3PI/2), ...)

(NaN,+-Inf) CX None

1448 TANH(CDX) |DX2|>=3.53D+15 (NaN,NaN) DX2 None

1449 TANH(CDX) CDX is close tosingularity ((0,+-PI/2),(0,+-3PI/2), ...)

(NaN,+-Inf) CDX None

1450 TANH(CQX) |QX2|>=2.0**62*PI

(NaN,NaN) QX2 None

- 196 -

Errornumber



data1 data2

1451 TANH(CQX) CQX is close tosingularity ((0,+-PI/2),(0,+-3PI/2), ...)

(NaN,+-Inf) CQX None

1452 ATANH(CX) CX=(+-1.0,0.0) (+-Inf,0.0) CX None

1453 ATANH(CDX) CDX=(+-1.0,0.0) (+-Inf,0.0) CDX None

1454 ATANH(CQX) CQX=(+-1.0,0.0) (+-Inf,0.0) CQX None

1455 DSHIFTL(I, J, SHIFT)

SHIFT<0 or

SHIFT>BIT_SIZE(I)

0 is set to SHIFT in SHIFT<0or BIT_SIZE(I) is set toSHIFT inSHIFT>BIT_SIZE(I), andprocess is continued

None

DSHIFTR(I, J, SHIFT)

SHIFTA(I, SHIFT)

SHIFTL (I, SHIFT)

SHIFTR (I, SHIFT)

MASKL(I [, KIND])I<0 or

I>BIT_SIZE(I)

0 is set to I in I<0 orBIT_SIZE(I) is set to I inI>BIT_SIZE(I), and processis continued

MASKR(I [, KIND])

- C1, C2, and C3 are one-byte character data. The value 26(X'1A') is assigned to C3.

- IX, IX1, and IX2 are four-byte integer data.

- JX, JX1, and JX2 are eight-byte integer data.

- X, X1, and X2 are real data.

- DX, DX1, and DX2 are double-precision real data.

- QX, QX1, and QX2 are quadruple-precision real data.

- CX is complex data. CX indicates (X1,X2).

- CDX is double-precision complex data. CDX indicates (DX1,DX2).

- CQX is quadruple-precision complex data. CQX indicates (QX1,QX2).

- PI is about 3.141592.

Note:

In many case, the arguments passed to the user-defined error subroutines are same as or part of argument of intrinsic procedurereferencing, however it is not same area.

8.1.6 Intrinsic Subroutine Error ProcessingThe table below lists the causes of intrinsic subroutine errors and the standard system correction processing.

The following arguments are passed to the user-defined error subroutine:

- Return code

- Error number

- The data listed in the table bellow

- 197 -

Table 8.4 Error processing for intrinsic subroutineError

numberReference format Cause of error Standard correction

processingData passed to

user-definedsubroutine

1301 MVBITS(FROM,FROMPOS,LEN,TO,TOPOS)

LEN < 0

or

LEN >BIT_SIZE(FROM)

TO is set to Zero None

FROMPOS < 0

or

FROMPOS >=BIT_SIZE(FROM)-LEN


TOPOS < 0

or

TOPOS >=BIT_SIZE(FROM)-LEN


1302 DATE_AND_TIME([DATE,TIME,ZONE,VALES])

The length ofDATE is less than8.

The value to be returned istruncated from the samelength as DATE.

None

The length ofTIME is less than10.

The value to be returned istruncated from the samelength as TIME.

None

The length ofZONE is less than5.

The value to be returned istruncated from the samelength as ZONE.

None

1303 DATE_AND_TIME([DATE,TIME,ZONE,VALES])

Size of VALES isless seven

The values are returned inVALES, and the rest of thevalues to be returned arediscarded.

None

1304 RANDOM_SEED([SIZE,PUT,GET])

Size of PUT is lessSIZE

The seed values are returnedin PUT, and therest of thevalues to be returned arediscarded.

None

Size of GET is lessSIZE

The seed values are returnedin GET, and therest of thevalues to be returned arediscarded.

None

1305 RANDOM_SEED([SIZE,PUT,GET])

Two or threearguments arespecified.


1306 MOVE_ALLOC([FROM,TO]) The length ofargument FROMand TO of thecharacter type isdifferent


- 198 -

Errornumber



1422 IEEE_GET_HALTING_MODE(FLAG,HALTING)

The function usedis not supported.

Nothing is done. None

IEEE_SET_HALTING_MODE(FLAG,HALTING)

IEEE_GET_UNDERFLOW_MODE(GRADUAL)

IEEE_SET_UNDERFLOW_MODE(GRADUAL)

IEEE_VALUE(X,CLASS)

1423 IEEE_SET_ROUNDING_MODE(ROUND_VALUE)

The function usedis not supported.

Nothing is done. None

8.1.7 Exception Handling ProcessingThis system has several program interrupts for error processing. These interrupts are exponent overflow, exponent underflow, divisioncheck, and integer overflow. If runtime option -i is specified, no exception handling is taken. See Section "3.3 Runtime Options" andSection "8.2.2 Debugging Programs for Abend" for details.

The table below shows an exception handling.

Table 8.5 Exception handling processing

Errornumber

Interrupt cause Standardcorrection

User-defined correction Data passed tothe user-definederror subroutine

Data1 Data2

11 The result of the floating- pointoperation exceeds |3.40282347e+38|(real) |1.797693134862316d+308| (double-precision real)

Terminates theprogram

Not correctable None

12 The result of the floating- pointoperation is less than |1.17549435e-38|(real) |2.225073858507201d- 308|(double-precision real)



13 The divisor in floatingpoint divisionis zero.



14 The divisor in fixedpoint division iszero, or absolute value of quotient ismore than 2**31.



17 Terminated abnormally due to runtime option -t



18 Terminated abnormally due to runtime option -a



19 Terminated abnormally due to aninvalid access.



20 Detected an error duringabnormally termination process



- 199 -

Errornumber

Interrupt cause Standardcorrection

User-defined correction Data passed tothe user-definederror subroutine

Data1 Data2

292 An invalid calculation in floating-point is executed.



Note:

- When standard correction processing sets a value in the result register, the sign is unchanged. In addition, the maximum values areas follows:Default real : 3.40282347e+38Double-precision real : 1.797693134862316d+308

- The above interrupts do not occur for quadruple-precision real. See Section "8.1.5 Intrinsic Function Error Processing" for details.

8.1.8 Other Error ProcessingThe table below lists the error processing performed when the system detects an error other than an input/output, intrinsic function, orprogram interrupts error. The following arguments are passed to the user-defined error subroutine:

- Return code

- Error number


Table 8.6 Other error processing

Errornumber

Standard correction processing User-defined correction Data passed to theuser-defined error

subroutine

301 Continues processing, assuming that first argument isdefault value


304 CALL PRNSET(IX)

Continues processing using IX = 0 for IX < 0, or IX =15 for IX > 15

Not correctable IX

306 Terminates the program Not correctable None


311 to318

Continues checking arguments Not correctable None

320 Continues processing Not correctable None

322 to324

Continues processing Not correctable None




1003 to1008

Terminates the program Not correctable None


1032 The multiple number is defined by CPU number andthe program is continued.


1033(*1)


- 200 -

Errornumber

Standard correction processing User-defined correction Data passed to theuser-defined error

subroutine




1038 to1041


1042 The multiple number is defined by num and theprogram is continued.



1044 to1046






1071 to1072


1141 to1143






1563 to1564


1565 to1566



1570 to1571

Continues checking arguments Not correctable None

1572 to1575


1576 to1577






IX: Default integer scalar

*1: The diagnostic message (jwe1033i-i) indicates that the Fortran program has been terminated during execution on multipleprocessors. If this message is output, the results of the Fortran program may be different from the results of terminating the programwhile running on a single processor by executing a STOP/ERROR STOP statement or an EXIT service subroutine, or by exceedingthe error count limit.

- 201 -

8.2 Debugging FunctionsThis section describes the check functions for debugging Fortran programs and the debug information output when execution endsabnormally.

8.2.1 Debugging Check FunctionsThe debugging check functions perform debugging during compilation or execution.

The -Ncheck_global compiler option is available as check functions to perform debugging during compilation.

During compilation, the compiler option -Ncheck_global checks the following:

- The procedure characteristics between external procedure definitions and references

- The procedure characteristics between external procedure definitions and interface body.

- In program units the size of common blocks

During compilation, the compiler option -Ha or -Nquickdbg=argchk checks the following:

- Array size in procedure in explicit reference specification.

During compilation, the compiler option -Hs or -Nquickdbg=subchk checks the following:

- The ranges of the declaration and reference for array element, and substring.

During execution, the compile option -H and -Nquickdbg check the following:

Table 8.7 List of runtime check items

-H option -Nquickdbg option

Checking Argument Validity (ARGCHK)

Checking the Number of Arguments

-Ha

-Nquickdbg=argchkChecking the Argument Types

Checking Argument Attributes -

Checking Function Types -Nquickdbg=argchk

Checking Argument Sizes -

Checking Explicit Interface -Nquickdbg=argchk

Checking Rank of Assumed-Shape Array -

Checking Subscript and Substring Values (SUBCHK)

Checking Array and Substring References-Hs

-Nquickdbg =subchk

Checking Array Section References -

Checking References for Undefined Data Items (UNDEF)

Checking for Undefined Data

-Hu

-Nquickdbg =undef

-Nquickdbg =undefnan

Checking for an Allocatable Variable -

Checking for an Optional Argument -

Shape Conformance Check (SHAPECHK)

Shape Conformance Check -He -

Extended Checking of UNDEF (EXTCHK)

Checking for Undefined Data of a Module and a CommonBlock -Hx

-

Checking for Unassociated Pointer -

- 202 -


Checking for Stride and Increment Value -

Checking for DO Variable -

Overlapping Dummy Arguments and Extend Undefined CHECK (OVERLAP)

Overlapping Dummy Arguments

-Ho

-

Checking for undefined Assumed-size Array withINTENT(OUT)

-

Checking for undefined variable of SAVE attribute -

Simultaneous OPEN and I/O Recursive Call CHECK (I/O OPEN)

Checking Connection of a File to A Unit-Hf

-

I/O Recursive Call CHECK -

- : No support

Details of checked items are given in the following sections:

- Check argument validity. (See Section "8.2.1.1 Checking Argument Validity (ARGCHK)")

- Check subscript and substring values. (See Section "8.2.1.2 Checking Subscript and Substring Values (SUBCHK)")

- Check references for undefined data. (See Section "8.2.1.3 Checking References for Undefined Data Items (UNDEF)")

- Check shape conformance. (See Section "8.2.1.4 Shape Conformance Check (SHAPECHK)")

- Extended check references for undefined data. (e.g. the variable declared in a module or in a common block).(See Section "8.2.1.5Extended Checking of UNDEF (EXTCHK)")

- Check the part of overlap is changed when two dummy arguments are overlapped. (See Section "8.2.1.6 Overlapping DummyArguments and Extend Undefined CHECK (OVERLAP)")

- Check the file is connected with two devices or more. (See Section "8.2.1.7 Simultaneous OPEN and I/O Recursive Call CHECK (I/O OPEN)")

- Check an input/output statement is executed from other input/output statement. (See Section "8.2.1.7 Simultaneous OPEN and I/ORecursive Call CHECK (I/O OPEN)")

The compiler option -H and -Nquickdbg check have the characteristics shown following:

Table 8.8 Characteristics of runtime check functions


Effect on execution performance Large Small

Checked items Many Few

Thread parallelism (*1) Not possible Possible

*1: Checks are possible/not possible during OpenMP or automatic parallel execution.

Compared with the check function of the compiler option -H, the check function of the compiler option -Nquickdbg has less effect onexecution performance because the items checked are limited. Debugging is also enabled during OpenMP and automatic parallel execution.

The compiler option -Nquickdbg includes the compiler option -Nquickdbg=inf_detail and -Nquickdbg=inf_simple.

If the compiler option -Nquickdbg=inf_detail is enabled, in addition to the message and line number where the error occurred, informationis output such as the variable name, procedure name, and element location in order to identify the cause of the error.

If the compiler option -Nquickdbg=inf_simple is enabled, the error and the line number where it occurred are displayed in the diagnosticmessage. However, if the compiler option -Nnoline is enabled, the value of line number is not guaranteed.

By limiting the information included in diagnostic messages, the compiler option -Nquickdbg=inf_simple has less impact on executionperformance than the compiler option -Nquickdbg=inf_detail check function.

See Section "2.2 Compiler Options" for more information about -H, -Nquickdbg, and -Ncheck_global.

- 203 -

8.2.1.1 Checking Argument Validity (ARGCHK)When the subprogram, a subroutine or function, is called, the compiler option -Ha or -Nquickdbg=argchk checks the validity of the dummyand actual arguments.

However, checks are not performed in the following cases:

- The actual argument is an actual name that has a POINTER attribute (including cases in which the actual argument is a derived typevariable with one or more components that have a POINTER attribute).

- An actual argument is an array section or array expression (other than an array variable).

- A dummy procedure name is used to call a subprogram.

- A derived type actual argument with an ultimate component that is an allocatable variable.

- A procedure pointer or type bound procedure is referred.

8.2.1.1.1 Checking the Number of Arguments

The compiler option -Ha or -Nquickdbg=argchk checks whether or not the number of actual arguments matches the number of dummyarguments. The number of actual arguments also includes the alternate return specifier. If the numbers of arguments do not match, adiagnostic message jwe0313i-w is output.

8.2.1.1.2 Checking the Argument Types

The compiler option -Ha or -Nquickdbg=argchk checks whether or not the dummy argument type matches the corresponding actualargument type. If the argument types do not match, a diagnostic message jwe0315i-w is output.

Argument types are not checked in the following cases:

- The dummy argument is a procedure name or an asterisk (*).

- The actual argument is an external procedure name or an alternate return specifier (but argument types are checked if the dummyargument and actual arguments are external procedure names that have types).

If the compiler option -Ha is set, the following checks are also performed

If the actual argument and dummy argument are derived type variables, the function checks whether or not the types of the componentsmatch. If the types do not match, a diagnostic message jwe1561i-w is output.

If the argument type is the character type, the function checks whether or not the length of the dummy argument exceeds that of the actualargument.

If the length is exceeded, a diagnostic message jwe0317i-w is output.

8.2.1.1.3 Checking Argument Attributes

The compiler option -Ha checks whether or not the dummy argument attributes matches the corresponding actual argument attributes. Adummy argument may be associated only with an actual argument as shown in "Table 8.9 Correspondence for dummy and actualarguments". If a dummy argument is not associated, the diagnostic message jwe0314i-w is output. If a dummy argument has INTENT(IN),the actual argument must be defined.

If an actual argument is a constant or an expression that is not a variable, diagnostic message jwe0318i-w is output if the correspondingdummy argument is changed in the subprogram.

"Table 8.9 Correspondence for dummy and actual arguments" lists the correspondence between dummy and actual arguments.

Table 8.9 Correspondence for dummy and actual arguments

Dummy argument Actual argument

Scalar variable Scalar variable, substring, array element, constant, result of scalar expression

Array excluding character type Actual argument excluding character type Whole array, array section, arrayelement, substring of array element, result of array expression

- 204 -

Dummy argument Actual argument

Character type array Actual argument of character type Scalar variable, substring, constant, resultof scalar expression, whole array, array section, array element, substring ofarray element, result of array expression

Dummy procedure External procedure, intrinsic procedure, or module procedure

* Statement label

8.2.1.1.4 Checking Function Types

The compiler option -Ha or -Nquickdbg=argchk checks whether or not the type of function defined in a function subprogram matches thetype of the reference source external procedure. If the function types do not match, a diagnostic message jwe0311i-w is generated if thefunction types do not match.

If a function subprogram is referenced as a subroutine subprogram, or if a subroutine subprogram is referenced as a function subprogram,a diagnostic message jwe0330i-w is output.

If the compiler option -Ha is set, the following checks are also performed.

If the function type is the character type, the function checks whether or not the lengths are the same. If the function lengths are not thesame, a diagnostic message jwe0312i-w is output.

8.2.1.1.5 Checking Argument Sizes

For an array dummy argument, the compiler option -Ha checks whether or not the size of array dummy argument exceed the correspondingactual argument. If the upper bound of the final dimension of the dummy array is an asterisk (*), or the upper and lower bounds of thedummy array are 1, the size of the dummy array is assumed to be equal to the corresponding actual argument size. Then this check isn’tperformed.

If type of assumed-shape array is character type, character lengths are checked to confirm that the length of the dummy argument and theactual argument match. If the length does not match a diagnostic message jwe1571i-w is output.

As listed in "Table 8.10 Array size", the size of the actual argument that corresponds to a dummy argument array is obtained from theactual argument attribute. If the size of the dummy array exceeds the actual argument size, a diagnostic message jwe0316i-w or jwe1577i-w is output.

Table 8.10 Array size

Actual argument attribute Actual argument size

Array Whole array size

Array element Size from the specified array element up to the end of the array

Substring of array element Size from the beginning of the specified array element substring to theend of the array

Scalar expression of character type excludingarray element and substring of array element

Length for scalar expression of character type

8.2.1.1.6 Checking Explicit Interface

If an explicit reference specification is not present in an external procedure reference that requires an explicit reference specification whenthe compiler option -Ha or -Nquickdbg=argchk is set, a diagnostic message jwe1569i-w is output.

8.2.1.1.7 Checking Rank of Assumed-Shape Array

If a dummy argument is an assumed-shape array, when the compiler option -Ha is set, the rank is compare to actual argument and dummyargument. If the rank is different between actual argument and dummy argument, a diagnostic message jwe1570i-w is output.

8.2.1.2 Checking Subscript and Substring Values (SUBCHK)If the compiler option -Hs or -Nquickdbg=subchk is set, the validity of the referenced subscript or substring expression is checked againstthe declared subscript and substring expression.

- 205 -

However, this check is not performed in the following cases:

- The referenced expression has the POINTER attribute (including cases in which a structure component has a POINTER attribute).

- An assumed-shape array is referenced.

- An array section specified by a vector subscript is referenced.

- An array in a masked array assignment is referenced.

- A named constant is referenced.

- The parent string of substring is a scalar constant.

- Data is referenced by a declaration expression.

- A derived type variable with an ultimate component that is an allocatable variable.

- The subscript range of io-implied-do (but checked if the -Hs compiler option is set).

This check is performed during compilation and execution. If an error is detected during compilation, a w level diagnostic message isoutput. The items checked during execution are explained below.

8.2.1.2.1 Checking Array and Substring References

If the compiler option -Hs or -Nquickdbg=subchk is set, when substring, array element, or array section is referenced, the array referencecheck determines whether or not those subscript and substring ranges are within the declared ranges. If it isn’t within the declared range,a diagnostic message (-Hs, -Nquickdbg=inf_detail:jwe0320i-w, -Nquickdbg=inf_simple:jwe1606i-u) is output.

If the upper and lower bounds of the dummy array are 1, the dummy array is regarded as assumed-size array. If the diagnostic message isoutput by the compiler option -Nquickdbg=subchk, the maximum number of 8 byte integer is output as the upper bound of the last dimensionof assumed-size array.

If the compiler option -Hs is set, the following checks are also performed

If the upper limit of an explicit shape array is a smaller value than the lower limit, a diagnostic message jwe1566i-w is output.

If the compiler option -Hs and -Ha are set simultaneously, this check whether or not an array element or substring is within the correspondingactual argument range when it corresponds to a dummy array in which the upper limit of the last dimension is an asterisk (*) or in whichthe upper and lower limit is the constant 1. If it is not within the actual argument range, a diagnostic message (jwe0322i-w) is output.

8.2.1.2.2 Checking Array Section References

If the compiler option -Hs is set, the following check is performed.

If an array section with a subscript triplet is referenced, the result of adding the first subscript to the extent of the subscript triplet is checkedto determine whether the value exceeds the maximum value of an integer or is less than the minimum value of an integer. In either case,diagnostic message jwe1565i-w is output.

8.2.1.3 Checking References for Undefined Data Items (UNDEF)If the compiler option -Hu, -Nquickdbg=undef, or -Nquickdbg=undefnan is set, the function checks whether or not referenced data isdefined.

However, this check is not performed in the following cases:

- An actual name that has a POINTER attribute is referenced (including cases in which a structure component has a POINTER attribute,but checked if the compiler option -Nquickdbg=undef or -Nquickdbg=undefnan is set).).

- A structure component has an ALLOCATABLE attribute.

- An assumed-shape array is referenced (but checked if the compiler option -Nquickdbg=undef or -Nquickdbg=undefnan is set).

- An array section specified by a vector subscript is referenced.

- Data in a masked array assignment is referenced.

- A derived type variable that has an allocatable array to an ultimate component is referenced.

- A variable of derived type having length type parameter is referenced.

- 206 -

- An array allocatable to an ultimate component by a structure component is referenced (but checked if the compiler option -Nquickdbg=undef or -Nquickdbg=undefnan is set).

- A structure component is referenced (but checked if the compiler option -Hu or -Nquickdbg=undef is set).

- The entities that share a storage sequence have different types (but checked if the compiler option -Hu or -Nquickdbg=undef is set).

- The subscript range of io-implied-do (but checked if the compiler option -Hu is set).

If diagnostic messages jwe0320i-w, jwe0322i-w, jwe1565i-w, and jwe1566i-w are detected by the SUBCHK function checks when thecompiler option -Hu and -Hs are set simultaneously, the UNDEF function is disabled.

Undefined data checks are explained below.

8.2.1.3.1 Checking for Undefined Data

If the compiler option -Hu or -Nquickdbg=undef is set, when module declaration part or variable not included in common block isreferenced, the values is checked to be defined for that data. When array variable is referenced, check is performed for each array element.If derived type variable is referenced when the compiler option -Hu is set, check is performed for each of the structure variable components.

If values is not defined, a diagnostic message (-Hu,-Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u) is output.

If a function subprogram ends when the compiler option -Hu is set, the values in the function results is checked. If it is not defined in thefunction results, a diagnostic message (-Hu, -Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u) is output. The linenumber at that time becomes the END statement of the function subprogram.

If data value has the values below, a diagnostic message (-Hu, -Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u)is output, but, as the Fortran program is correct, no correction is required. To change the data values below, specify theFLIB_UNDEF_VALUE environment variable. See Section "3.8 Using Environment Variables for Execution" for detail.

One-byte integer : -117

Two-byte integer : -29813

Four-byte integer : -1953789045

Eight-byte integer : -8391460049216894069

Single-precision real : -5.37508134e-32

Double precision real : -4.696323204354320d-253

Quadruple precision real : -9.0818487627532284154072898964213742q-4043

Single-precision complex : (-5.37508134e-32,-5.37508134e-32)

Double precision complex : (-4.696323204354320d-253,-4.696323204354320d-253)

Quadruple precision complex : (-9.0818487627532284154072898964213742q-4043,

-90818487627532284154072898964213742q-4043)

Character : Z'8B'

8.2.1.3.2 Checking for Undefined Data Using Invalid Operation Exception

If the compiler option -Nquickdbg=undefnan is set, when module declaration part or variable not included in common block is referenced,the value is checked to be defined for that data. When array variable is referenced, check is performed for each array element. If value isnot defined, a diagnostic message (-Hu,-Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u) is output.

Output diagnostic message is different according to a variable and a type of function results.

If integer type and character type checks show that values are not defined, a diagnostic message (-Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u) is output.

If data has the values below, a diagnostic message (-Hu, -Nquickdbg=inf_detail:jwe0323i-w, -Nquickdbg=inf_simple:jwe1606i-u) isoutput, but, as the Fortran program is correct, no correction is required. To change the data values below, specify the environment variableFLIB_UNDEF_VALUE. See Section "3.8 Using Environment Variables for Execution" for detail.

One-byte integer : -117

Two-byte integer : -29813

- 207 -

Four-byte integer : -1953789045

Eight-byte integer : -8391460049216894069

Character : Z'8B'

The real type and complex type variable have SNaN as the initial value. If values aren’t defined, the real type and complex type checksdetect an invalid operation exception (jwe0292i-u).

However, the invalid operation exception may not be detected in the following case:

- The real type and complex type variables do not appear in the operands of the command that issues invalid operation exceptions. (TheCPU architecture determines which instruction issues invalid operation exception. For example, in an architecture where the referenceinstruction does not issue exception, undefined variable is detected during referencing when the compiler option -Nquickdbg=undefis set, but undefined variable is not detected during referencing when the compiler option -Nquickdbg=undefnan is set.)

- The invalid operation exception is masked.

In addition undefined data reference, the invalid operation exception is detected when the execution result is SNaN.

8.2.1.3.3 Checking for an Allocatable Variable

If the compiler option -Hu is set, the variable checked to already allocated by ALLOCATE statement when an ALLOCATABLE variableis referenced; if not, diagnostic message jwe0324i-w is generated.

8.2.1.3.4 Checking for an Optional Argument

If the compiler option -Hu is set, the argument checked to be present when an optional argument is referenced; if not, diagnostic messagejwe1575i-s is generated.

8.2.1.4 Shape Conformance Check (SHAPECHK)When an array expression or array assignment statement is executed, the compiler option -He checks shape conformance.

Diagnostic message jwe0329i-w is generated if a mismatch of shape conformance occurs during execution.

8.2.1.5 Extended Checking of UNDEF (EXTCHK)The -Hx option checks to see if the variable declared in a module or in a common block is defined when the variable is referenced, to seeif the pointer is associated when a pointer is referenced, and to see if the incrementation parameter of the DO construct, stride of theFORALL, incrementation parameter of array constructor implied do control and stride of subscript-triplet is not 0.

Diagnostic message jwe0323i-w is generated for the variable declared in a module or in a common block that is not defined. Note that themessage is generated if the item is defined with a value listed in Section "8.2.1.3 Checking References for Undefined Data Items (UNDEF)", even though the Fortran program may be correct.

A Pointer is referenced, the pointer checked to be associated with target; if not, diagnostic message jwe1572i-s is generated. However,this check facility is not performed for a pointer of derived type and a pointer of component.

If the incrementation parameter of the DO construct, stride of the FORALL, incrementation parameter of array constructor implied docontrol and stride of subscript-triplet is zero, diagnostic message jwe1573i-s is generated.

In the range of a DO construct, if an actual argument is a do-variable, the definition of do-variable is checked. Diagnostic message jwe1574i-s is generated if the do-variable is defined.

The execution of a RETURN or an END statement within a subprogram causes all local variables in that scope unit without the SAVEattribute becomes undefined when the -Hx option is present.

When the variables have a BIND attribute, the undefined value is not checked.

If the -Hx option is specified, the -Hu option is in effect, see Section "8.2.1.3 Checking References for Undefined Data Items (UNDEF)".

Be careful to the following when the -Hx option using:

- All program units that have a definition or initialization for common block object shall be compiled with -Hx option.

Example: Invalid specification of -Hx

File: a.f90

- 208 -

BLOCK DATA INITCOMMON /CMN/ JDATA J/200/END BLOCK DATA

File: b.f90

PROGRAM MAINCOMMON /CMN/ JPRINT *,JEND PROGRAM

Command:

$ frtpx a.f90 -c$ frtpx b.f90 a.o -Hx

The -Hx options shall be specified for compiling a.f90.

- All program units that uses the module shall be compiled with -Hx option, if the module is compiled with -Hx option.

Example: Invalid specification of -Hx

File: a.f90

MODULE MODINTEGER :: JEND MODULE

File: b.f90

PROGRAM MAINUSE MODJ = 200CALL SUB()END PROGRAM

File : c.f90

SUBROUTINE SUB()USE MODPRINT *,JEND SUBROUTINE

Command:

$ frtpx a.f90 -c -Hx$ frtpx b.f90 -c$ frtpx c.f90 a.o b.o -Hx

The -Hx options shall be specified for compiling b.f90.

8.2.1.6 Overlapping Dummy Arguments and Extend Undefined CHECK (OVERLAP)When compiler option -Ho is specified, the following check is done.

- Two dummy arguments are overlapped and the part of overlap is changed.

- When an assumed-size array with INTENT(OUT) attribute is referenced during execution, checks to see if the variable is defined.

- When a variable with SAVE attribute is referenced during execution, checks to see if the variable is defined.

8.2.1.6.1 Overlapping Dummy Arguments

Do not define or undefine the value of the part overlapping executing the procedure when two dummy arguments are overlapping in thesame procedure. When compiler option -Ho is specified, the function checks whether the value is defined or undefined while executing.

- 209 -

Diagnostic message jwe1576i-w is generated if a mismatch occurs during execution. The validity is not checked if any of the followingconditions is true.

The validity is not checked if:

- The dummy argument is defined in a masked array assignment.

- The dummy argument is defined in FORALL assignment statement.

- The dummy argument as DO variable is defined.

- The dummy argument is defined with array section of vector subscript.

- The dummy argument is defined and the corresponding actual argument has no continuation element.

- The dummy argument is defined in reference of procedure.

- When the dummy argument which has TARGET attribute or POINTER attribute is overlapping the object of the other dummy argumentwhich has neither TARGET attribute nor POINTER attribute, and the dummy argument is defined.

- When the dummy argument of assumed shape array is overlapping the object of the other dummy argument, and the dummy argumentis defined.

8.2.1.6.2 Undefined Assumed-size Array with INTENT(OUT)

When compiler option -Ho is specified, the function checks whether assumed-size array with INTENT(OUT) attribute is defined whileexecuting. Diagnostic message jwe0323i-w is generated if a mismatch occurs during execution. If compiler option -Hs also is specified,the function does not check the data if diagnostic message jwe0320i-w, jwe0322i-w, jwe1565i-w, or jwe1566i-w is generated.


- A pointer variable is referenced.

- The referenced expression is an array section with a vector subscript

- The referenced expression is in a masked array assignment.

- The referenced expression is a derived type variable with an ultimate component that is an allocatable variable.

- The referenced expression is an allocatable variable as an ultimate component by a structure component.

8.2.1.6.3 Undefined Variable of SAVE Attribute

When compiler option -Ho is specified, the function checks whether variable with SAVE attribute is defined while executing. Diagnosticmessage jwe0323i-w is generated if a mismatch occurs during execution. If compiler option -Hs also is specified, -Hu option does notcheck the data if diagnostic message jwe0320i-w, jwe0322i-w, jwe1565i-w, or jwe1566i-w is generated.


- The referenced expression has a structure variable one of whose structure components has the POINTER attribute.

- The referenced expression is an array section with a vector subscript.

- The referenced expression is in a masked array assignment.

- The referenced expression is a derived type variable with an ultimate component that is an allocatable variable.

- The referenced expression is an allocatable variable as an ultimate component by a structure component.

8.2.1.7 Simultaneous OPEN and I/O Recursive Call CHECK (I/O OPEN)When compiler option -Hf is specified, the following check is done.

A file is connected with two devices or more at the same time in the input/output statement.

An input/output statement is called by the function while executing other input/output statement.

- 210 -

8.2.1.7.1 Checking Connection of a File to A Unit

In the input/output statement, do not connect the one file with two devices or more at the same time. When compiler option -Hf is specified,the function checks whether the one file is connected with two devices or more at the same time while executing. Diagnostic messagejwe1231i-w is generated if a mismatch occurs during execution.


- The file is specified by the file names other than the real name when file system that the file is not local is used.

- As for this file, specification different from the real name is used when the file is defined and the alias is defined by the "alias"command.

- "../" is included in the path of the file.

8.2.1.7.2 I/O Recursive Call CHECK

Do not recurrently execute the input/output statement. When compiler option -Hf is specified, the function checks whether the input/outputstatement is executed recurrently while executing. Diagnostic message jwe1232i-s is generated if a mismatch occurs during execution.

8.2.2 Debugging Programs for AbendIf an abnormal termination event (hereafter called an abend) occurs during execution of a Fortran program, the system generates informationto help the user determine the cause of the abend.

8.2.2.1 Causes of Abend"Table 8.11 Detected errors and corresponding signal codes" lists the errors that the system detects and the corresponding signal codes.If the -i runtime option is specified, the system does not check for these errors.

If the -NRnotrap compiler option is specified, the system does not check for floating point exceptions (jwe0011i-u, jwe0012i-u, jwe0013i-u or jwe0292i-u). For more information about compiler option -NRnotrap, see Section "2.2 Compiler Options".

Table 8.11 Detected errors and corresponding signal codes

Signal number Meaning of signalnumber

Signal code Meaning of signal code

SIGILL(04) *Incorrect instructionexecuted

1 ILL_ILLOPC Illegal opcode

2 ILL_ILLOPN Illegal operand

3 ILL_ILLADR Illegal addressing mode

4 ILL_ILLTRP Illegal trap

5 ILL_PRVOPC Privileged opcode

6 ILL_PRVREG Privileged register

7 ILL_COPROC Coprocessor error

8 ILL_BADSTK Internal stack error

SIGFPE(08) Arithmetic exception

1 FPE__INTDIV * Fixed-point division exception

3 FPE__FLTDIV Floating-point division exception

4 FPE__FLTOVF Floating-point overflow exception

5 FPE__FLTUND Floating-point underflow exception (*1)

7 FPE_FLTINV Invalid floating-point processing

SIGBUS(10) *Storage protectionexception

1 BUS_ADRALN Invalid address alignment

2 BUS_ADRERR Non-existent physical address

3 BUS_OBJERR Object specific hardware error

SIGSEGV(11) * Segmentation exception 1 SEGV_MAPERR Address not mapped to object

- 211 -

Signal number Meaning of signalnumber

Signal code Meaning of signal code

2 SEGV_ACCERR Invalid permissions

SIGXCPU(30) * CPU time interrupt -- -- --

*1: Interrupt detected when the -NRtrap compiler option and the environment variable FLIB_EXCEPT=u (exponent underflow) werespecified.

Note:

1. For signal numbers marked with an asterisk, a core file is created when the -i runtime option is specified to disable system errordetection.

8.2.2.2 Information Generated when an Abend OccursDepending on the error type, information is written to the standard error file as explained in the following sections. Information that cannotbe written to the standard error file is sent to the terminal.

8.2.2.2.1 Information Generated for a General Abend

The information generated for a general abend is follows. This information is generated when SIGILL, SIGBUS, or SIGSEGV is detected.A trace back map is printed after this information.

jwe0019i-u The program was terminated abnormally with signal number SSSSSSS. signal identifier = NNNNNNNNNNN, (Detailed information.)

SSSSSSS: SIGILL, SIGBUS, or SIGSEGV

NNNNNNNNNNN: Signal code for the abend cause

(Detailed information.): Detailed information of Signal code NNNNNNNNNNN

When it is strong possibility that the cause of SIGBUS or SIGSEGV is satck-overflow, add the following information.

The cause of this exception may be stack-overflow

8.2.2.2.2 Information of SIGXCPU

The output when SIGXCPU is detected is follows.

jwe0017i-u The program was terminated with signal number SIGXCPU.

8.2.2.2.3 Information Generated by Run-Time Option -a

The information generated by runtime option -a when SIGIOT is detected is follows. The program is forced to abend without deleting thetemporary files being used by the program and a core file is created.

The cause of this exception may be stack-overflow.

jwe0018i-u The program was terminated abnormally due to runtime option -a with signal number SIGIOT.

8.2.2.2.4 Information when an Abend Occurs Again during Abend Processing

The output when an abend occurs again during abend processing is follows.

jwe0020i-u An error was detected during an abnormal termination process.

8.2.3 Hook FunctionThis section describes the hook function which calls a user-defined subroutine by specified location in a Fortran program, or by regulartime interval. Program operation can be checked via a user-defined subroutine to output trace information and specific variable values.

- 212 -

8.2.3.1 User-defined Subroutine FormatThe user-defined subroutine name is fixed (USER_DEFINED_PROC).

Define a user-defined subroutine in the format shown below.

Format:

SUBROUTINE USER_DEFINED_PROC(FLAG, NAME, LINE, THREAD)INTEGER(KIND=4),INTENT(IN):: FLAG, LINE, THREADCHARACTER(LEN=*),INTENT(IN):: NAME

Arguments:

FLAG : Indicate the calling source for the user-defined subroutine.

= 0 : Program entry

= 1 : Program exit

= 2 : Procedure entry

= 3 : Procedure exit

= 4 : Parallel region (OpenMP or automatic parallelization) entry

= 5 : Parallel region (OpenMP or automatic parallelization) exit

= 6 : Regular time interval

= 7~99 : Reserved for system

= 100~ : Can be used by the user

NAME : Indicate the calling source procedure name.

NAME can be referenced only if the FLAG is 2, 3, 4, 5, or 100 or greater.

LINE : Indicate the calling source line number.

LINE can be referenced only if the FLAG is 2, 3, 4, 5, or 100 or greater.

THREAD : With thread parallelization using OpenMP or automatic parallelization, indicate the number ofthe thread that calls the user-defined subroutine.

THREAD can be referenced only if the FLAG is 2, 3, 4, 5, or 100 or greater.

8.2.3.2 Notes on the Hook FunctionThe following notes apply when the hook function is used:

- The procedure name "USER_DEFINED_PROC" must not be used for any other purpose.

- Operation is not guaranteed if the user-defined subroutine is not defined.

- Values must not be set in the user-defined subroutine arguments.

- If the -Nnoline compile option is set, the LINE argument value is not guaranteed.

- In thread parallelized programs, the user-defined subroutine may be called from multiple threads.

- Operation is not guaranteed if the user-defined subroutine is compiled with the -CcI4I8 option set.

- Operation is not guaranteed if the user-defined subroutine is compiled with the -AU option set.

- In procedures that include an ENTRY statement, the procedure name set in the NAME argument may be the entry name of the ENTRYstatement.

- The procedure exit becomes the END statement line number.

- The characteristics of a user-defined subroutine procedure cannot be changed.

- The entry and exit of user-defined subroutine except the calling from any location in the program must not be calling source for theuser-defined subroutine.

- 213 -

8.2.3.3 Calling a User-defined Subroutine from a Specified LocationWhen -Nhook_func option is specified when the object program and executable program are created, user-defined subroutine can be calledfrom the following locations:

- Program entry and exit

- Procedure entry and exit

If the -Kopenmp or -Kparallel option is enabled, user-defined subroutine can also be called from the following locations:

- Parallel region (OpenMP or automatic parallelization) entry and exit

The following internal procedure name is passed to NAME argument in OpenMP, automatic parallelization entry and exit.

- OpenMP "_OMP_IdNumber_"

- automatic parallelization "_OMP_IdNumber_"

When the -Nhook_func is set, operation is as follows for Fortran, and C linked programs:

- If user-defined subroutine/function is defined within the Fortran program and the C program respectively, the user-defined subroutinedefined in the Fortran program is called from the Fortran program, and the user-defined function defined in the C program is calledfrom the C program.

- If user-defined subroutine/function is defined within only the Fortran program or only the C program, the defined user-definedsubroutine/function is called from both the Fortran program and the C program.

An example using the hook function with location specified is shown below.

User-defined subroutine program example:

File: hook.f95

WRITE(*,*) 'HELLO'CALL SUBEND

SUBROUTINE SUBWRITE(*,*) 'SUB'END

SUBROUTINE USER_DEFINED_PROC(FLAG,NAME,LINE,THREAD)INTEGER(KIND=4),INTENT(IN):: FLAG,LINE,THREADCHARACTER(LEN=*),INTENT(IN):: NAMESELECT CASE(FLAG) CASE(0) WRITE(*,*) 'PROGRAM START' CASE(1) WRITE(*,*) 'PROGRAM END' CASE(2) WRITE(*,*) 'PROC START: ',NAME,' LINE: ',LINE CASE(3) WRITE(*,*) 'PROC END: ',NAME,' LINE: ',LINEEND SELECTEND

Executable program creation:

$ frt -Nhook_func hook.f95

Execution results:

$ ./a.out PROGRAM START HELLO PROC START: SUB LINE: 5 SUB

- 214 -

PROC END: SUB LINE: 7 PROGRAM END$

8.2.3.3.1 Notes on Calling from a Specified Location

The following notes apply when the -Nhook_func option is set:

- If a procedure with few computations is called repeatedly, calling a user-defined subroutine from a procedure entry/exit may affectexecution performance.

- If a parallel region with few computations is executed repeatedly, calling a user-defined subroutine from a parallel region entry/exitmay affect execution performance.

- User-defined subroutine is not called from procedures called from input-output lists.

8.2.3.4 Calling a User-defined Function at Regular Time IntervalWhen the -Nhook_time option is set when an executable program is created, the user-defined function is called by regular user timeinterval.

The calling interval can be specified using the FLIB_HOOK_TIME environment variable. If it is not specified, the user-defined functionis called once every minute. See Section "3.8 Using Environment Variables for Execution", for information about the FLIB_HOOK_TIMEenvironment variable.

See the "C User's Guide" or the "C++ User's Guide", for information about the user-defined function.

8.2.3.4.1 Notes on Calling by Regular Time interval

The following notes apply when the -Nhook_time option is set.

- The environment variable FLIB_HOOK_TIME specification must not be changed while program execution is in progress.

- Operation is not guaranteed if a value outside the range 0 to 2147483647 is specified in the environment variable FLIB_HOOK_TIME.

- If the calling interval is short, the user-defined function is called at the regular time intervals but execution performance may beaffected.

- The user-defined function must be defined in C or C++ program.

This function calls the user-defined function from asynchronous signals. Therefore, the following restrictions apply to calling the user-defined function at regular time interval:

- Functions excluding async-signal-safe functions cannot be used in the user-defined function.

- When the -Nhook_time option is set, signal processing routines corresponding to SIGVTALRM must not be defined.

- When the -Nhook_time option is set, ALARM(3) service function must not be used.

- The debugger and profiler must not be used for the executable program created with the -Nhook_time option.

8.2.3.5 Calling from Any Location in the ProgramA user-defined subroutine can be called from any location in a program by specifying arguments (FLAG:100~, NAME, LINE, THREAD).

- 215 -

Chapter 9 Optimization FunctionsThis chapter explains functions to use optimization functions effectively, as well as some points to note.

9.1 Overview of OptimizationOptimization aims to generate object modules (instructions and data areas) that allow the program to execute as quickly as possible.

Some optimization functions can significantly increase the size of the object program. Therefore, the amount of virtual storage space canalso increase.

Therefore, it can be selected whether to use optimization function by options.

See Section 2.2 Compiler Options for more information about optimizations parameters.

9.1.1 Standard OptimizationWhen an optimization level between level 1 (-O1 option) and level 3 (-O3 option) is specified, standard optimization is performed.

To determine whether each optimization was applied, specify the -Koptmsg=2 option.

9.1.1.1 Elimination of Common ExpressionIf two expressions that give equal calculation results (common expressions) are present, the result of the former expression can be usedin the latter expression without calculation.

An example of elimination of common expressions is given below.

Example:

[Original source-code] [Optimized pseudo-code]

...A=X*Y+C...B=X*Y+D

->

T=X*YA=T+C...B=T+D

The X*Y portion of the expression on the right hand side of the assignment is a common expression. The code is changed so that thesecond calculation is eliminated and the result of the first calculation T is used instead. T is a variable generated by the compiler.

9.1.1.2 Movement of Invariant ExpressionsExpressions with values that do not change within a loop (invariant expressions) are moved outside the loop.

An example of the movement of invariant expressions is given below.

Example:


DO I=1,N Y=A(J)*2 X=X+Y*ZENDDO

->

Y=A(J)*2T=Y*ZDO I=1,N X=X+T

ENDDO

The entire statement Y=A(J)*2 and the Y*Z portion of the expression are moved outside the loop. T is a variable generated by thecompiler.

The objects that are usually moved by this optimization are statements or parts of statements that are always executed within the loop.

However, if the -Kpreex option is specified, invariant expressions in statements that are selectively executed within the loop dependingon a decision statement are also moved.

- 216 -

To differentiate this optimization from the normal movement of invariant expressions, it is called advance evaluation of invariantexpressions. This optimization can further reduce execution time.

However, side effects may occur due to the movement. For details of the side effects, see Section 9.1.2.3 Advance evaluation of invariantexpressions.

9.1.1.3 Reducing the Strength of OperatorsThe "strength" of operators describes the relative amount of execution time required for operation. A "strong" operator requires a largeamount of execution time. Typically, addition and subtraction have the same strength. Multiplication is stronger than addition or subtractionand division is even stronger than multiplication. In addition, type conversion between the integer and float types is stronger than additionor subtraction but weaker than multiplication.

An example of reducing strength is given below.

Example:


DO I=1,10 ... J=K+I*100 ... ...ENDDO

->

T=100DO I=1,10 ... J=K+T ... T=T+100ENDDO

Optimization is performed on the operation I*100 on the derivative variable I, so the operation I*100 within the loop is converted tothe addition T=T+100. t is a variable generated by the compiler.

9.1.1.4 Loop UnrollingApplying loop unrolling, all of the executable statements in a loop are unrolled n times and the loop iteration count is divided by n.However, if the compile-time option -Ksimd[=level] is set and the loop has been SIMD optimized, the executable statement is unrolled2*n, and the loop iteration count is divided by 2*n.

The value of n is determined by the compiler, depending on the iteration count and the number of executable statements in the loop.

The object module size increases because the statements in the loop are unrolled many times.

Loop unrolling can be suppressed by specifying the compile-time option -Knounroll.

To output line numbers and expansion counts of optimized DO loops, specify -Koptmsg=2 option.

Statements that include array assignments (array expressions) are converted to DO loops, and will be subject to loop unrolling. However,the messages for -Koptmsg=2 are not output.

9.1.1.5 Loop BlockingLoop blocking is the optimization to subdivide the access for array by multiplexing the loop in block size. As a result, the localization ofdata access is improved, and then efficient use of the cache is promoted.

This optimization is performed on loops which contain no jump from the inside to the outside and no jump from the outside to the inside.

Loop blocking is performed in optimization level 2 (-O2 option). It can be suppressed by specifying the -Kloop_noblocking option.

The message to show the optimized loop can be output by specifying the -Koptmsg=2 option.

An example of loop blocking is given below.

Example:


DO K=1,NK DO J=1,NJ DO I=1,NI A(I,J)=A(I,J)+B(I,K)*C(K,J)

->

DO KK=1,NK,MK+1 DO JJ=1,NJ,MJ+1 DO II=1,NI,MI+1 DO K=KK,MIN(NK,II+MK)

- 217 -

ENDDO ENDDOENDDO

DO J=JJ,MIN(NJ,JJ+MJ) DO I=II,MIN(NI,II+MI) A(I,J)=A(I,J)+B(I,K)*C(K,J) ENDDO ENDDO ENDDO ENDDO ENDDOENDDO

9.1.1.6 Software PipeliningSoftware pipelining schedules instructions in a loop to execute as parallel as possible.

Software pipelining can be suppressed by specifying the compile-time option -Knoswp.

Software pipelining performs an instruction scheduling to arrange instructions of a particular iteration in a loop with instructions of thefollowing iterations, and reshapes the loop. Thus software pipelining requires enough loop iteration count. Other optimizations affect therequired iteration count because it depends on instructions in the loop. If the original loop is short, software pipelining may be less effectivebecause instruction scheduling is performed moderately to keep required iteration count small.

When software pipelining is applied to a loop whose iteration count is variable, as shown in the following example, the compiler insertsa branch instruction to choose the software-pipelined loop or the original loop in case the iteration count is short. It is decided at run timewhether the software-pipelined loop is chosen. The size of the object module will increase.

Example:

[Original source-code]

SUBROUTINE SUB(N)INTEGER(KIND=4) :: N(The original loop: the iteration count is N.)RETURNEND

[Optimized pseudo-code]

SUBROUTINE SUB(N)INTEGER(KIND=4) :: NIF (Is N enough?) THEN (The software-pipelined loop)ELSE (The original loop: the iteration count is N.)ENDIFRETURNEND

If the compile-time option -Koptmsg=2 is specified, optimization messages about results of software pipelining are output with the linenumbers of loops. When software pipelining is applied to a loop, optimization messages jwd8204o-i and jwd8205o-i are output at once.When software pipelining is not applied to a loop, one of optimization messages from jwd8661o-i to jwd8672o-i is output. Optimizationmessages about software pipelining are not output when the -Knoswp option is specified, or when a loop disappears because of otheroptimizations before software pipelining, such as full unrolling.

The optimization message jwd8205o-i shows the minimum iteration count required to choose the software-pipelined loop. Make sure toconsider following optimizations applied to the same loop because the target loop of software pipelining is the innermost loop; when loopcollapse, loop interchange, thread parallelization, or inline expansion is applied, the iteration count of the target loop of software pipeliningis changed as shown in the following table. If this is the case, the value shown in the optimization message jwd8205o-i should be comparedwith the iteration count in the table.

Optimization applied to

the same loopMessage number The iteration count of the target loop to which software

pipelining is applied

Loop collapse jwd8330o-i Product of the iteration counts of the collapsed loops.

- 218 -

Optimization applied tothe same loop

Message number The iteration count of the target loop to which softwarepipelining is applied

Loop interchange jwd8211o-i The iteration count of the innermost loop after loopinterchange.

Thread parallelization jwd5001p-i, etc. Quotient of the iteration count divided by the number ofthreads.

Inline expansion jwd8101o-i The iteration count of the innermost loop after inlineexpansion.

9.1.1.7 SIMD

9.1.1.7.1 Normal SIMD

The compiler may apply optimizations using SIMD extensions. Using SIMD extensions, multiple operations of the same kind are executedat once.

The optimizations using SIMD extensions are suppressed by specifying compiler option -Knosimd. If compiler option -Koptmsg=2 isspecified, messages listing the line numbers where optimized loops are placed are output.

9.1.1.7.2 SIMD Extension for Loop Containing IF Construct

When -Ksimd={2|auto} options is specified, SIMD extension can be applied to the loop that contains IF construct. The conditionalexecution instructions are only store instruction and move instruction in CPU architecture, and other instructions are executed speculatively(statements in the IF construct are executed even if the condition of the IF construct is false). Therefore when the true rate of condition ofthe IF construct is high, the performance may be improved.

However, side effects, such as exceptions at execution, may occur in the execution results because instructions are executed speculatively.

9.1.1.7.3 List Vector Conversion

When the IF construct in loop contains many instructions and the true rate of the IF construct is low, the performance may be improvedby applying list vector conversion.

The list vector conversion generates the following two loops:

1. The loop to save the value of loop control variable into a new array when the condition of IF construct is true.

2. The loop to execute the statement of the original IF construct, and its loop iteration count is equal to the number of the new arrayelement.

In the case of 2, that loop becomes list access method, and becomes the target of SIMD extension and software pipelining.

Example of list vector conversion is shown in the following. The list vector conversion is applied by specifying the optimization controlline. See Section "9.10.4 Optimization Control Specifiers" for information on "!OCL SIMD_LISTV[(ALL | THEN | ELSE)]".

Example:


DO I=1, N!OCL SIMD_LISTV IF (M(I)) THEN A(I)=B(I)+C(I) ENDIFENDDO


J=1DO I=1,N IF (M(I)) THEN IDX(J)=I J=J+1

- 219 -

ENDIFENDDO!OCL NORECURRENCEDO K=1,J-1 I=IDX(K) A(I)=B(I)+C(I)ENDDO

9.1.1.7.4 SIMD with Redundant Executions for the SIMD Width

This function is validated by the optimization control line "!OCL SIMD_REDUNDANT_VL(n)". This is applied to loops whose numberof iterations is equal to or less than n, which is the argument of the SIMD_REDUNDANT_VL specifier. When the number of iterationsof the loops is not a multiple of the SIMD width, SIMD instructions with masks are used to generate loops which use SIMD extensionsover the whole iterations. (Refer to the following example and figures.) Since instructions with masks are provided only for moves andstores, other instructions are executed redundantly.

The objective is to benefit by using SIMD extensions for loops whose number of iterations is small, while the following optimizations aresuppressed because they require plenty of iterations. Thus this function does not always improve performance.

- Software Pipelining

- Loop Striping

- XFILL

Note that the performance of loops such as the followings could decrease. Use this function for programs you understand well.

- Loops whose number of iterations tends to be a multiple of the SIMD width plus 0, 1, or 2

- Loops which use SIMD extensions to execute eight, which is double of the SIMD width, single-precision floating-point operationssimultaneously; the loops can be identified by the optimization information "SIMD (VL: 8)" output by means of the -Qt option or the-Nlst=t option.

The followings are an example and images of this function.

When the function is not applied to the source program in the example, SIMD instructions are used as shown in "Figure 9.1 The optimizationcontrol line is not effective (4-wide SIMD with M=15)". The remaining iterations, which correspond the remainder after division of thenumber of iterations by the SIMD width, do not use SIMD instructions.

When the function is applied, SIMD instructions are used as shown in "Figure 9.2 The optimization control line is effective (4-wide SIMDwith M=15 and the SIMD_REDUNDANT_VL(15) specifier)". The remaining iterations are executed by using SIMD instructions withmasks.

Example: The optimization control line (OCL) is specified

!OCL SIMD_REDUNDANT_VL(n) ! n is a value between 2 and 255 DO I=1,M A(I)=B(I)+C(I) END DO

- 220 -

Figure 9.1 The optimization control line is not effective (4-wide SIMD with M=15)

[Explanation]When the loop control variable I is from 13 to 15, SIMD instructions are not used.

Figure 9.2 The optimization control line is effective (4-wide SIMD with M=15 and the SIMD_REDUNDANT_VL(15)specifier)

[Explanation]The whole iterations are executed by using SIMD instructions with masks.

9.1.1.8 Loop UnswitchingLoop unswitching is an optimization to modify loop by putting out the IF construct out of the loop and creating a loop in each "TRUE"and "FALSE" blocks of the IF construct in order to promote some optimizations such as instruction scheduling if the condition of the IFconstruct is invariant in a loop.

The loop unswitching is enable by the -O3 option, and disabled by the -O2 option or less.


- 221 -


SUBROUTINE SUB(A,B,C,N)REAL*8 A(N)DO I=1,N IF (B .EQ. C) THEN A(I) = 0.0D0 ELSE A(I) = 1.0D0 ENDIFENDDOEND SUBROUTINE

->

SUBROUTINE SUB(A,B,C,N)REAL*8 A(N)IF (B .EQ. C) THEN DO I=1,N A(I) = 0.0D0 ENDDO ELSE DO I=1,N A(I) = 1.0D0 ENDDO ENDIFEND SUBROUTINE

9.1.1.9 Loop FissionLoop fission is the optimization to split a loop body into multiple small loops. This can improve locality of reference, both of the databeing accessed in the loop and the code in the small loop's body. As a result, other optimizations are promoted.

This optimization is performed on loops which contain no jump from the inside to the outside and no jump from the outside to the inside.

Loop fission is performed in optimization level 2 (-O2 option). It can be suppressed by specifying the -Kloop_nofission option.


An example of loop blocking is given below.

Example:


DO I=1,N E(I) = A1(I)*B1(I) + A2(I)*B2(I) + A3(I)*B3(I) + A4(I)*B4(I) + A5(I)*B5(I) F(I) = C1(I)*D1(I) + C2(I)*D2(I) + C3(I)*D3(I) + C4(I)*D4(I) + C5(I)*D5(I)ENDDO


DO I=1,N E(I) = A1(I)*B1(I) + A2(I)*B2(I) + A3(I)*B3(I) + A4(I)*B4(I) + A5(I)*B5(I)ENDDO

DO I=1,N F(I) = C1(I)*D1(I) + C2(I)*D2(I) + C3(I)*D3(I) + C4(I)*D4(I) + C5(I)*D5(I)ENDDO

9.1.2 Extended OptimizationWhen one of the extended function options is specified regardless of optimization, extended optimization is performed in addition to thestandard optimization.

This section describes the functions comprising extended optimization.

Depending on the extended optimization function, the size of the object modules (instructions and data areas) generated by the compilermay be greatly enlarged or side effects may occur in the execution results.

9.1.2.1 Inline ExpansionApplying inline expansion, user-defined procedures are inlined to referenced points.

The following expression shows how much size of object modules in program units increases when inline expansion is performed on user-defined procedures.

(Size of instructions + Size of data sections) * Expansion count

- 222 -

The size of instructions is determined by the kind and number of executable statements.

The size of data sections is the total size of variables, arrays, and constants, excluding common blocks, dummy arguments, allocatablevariables, and elements with host associations.

Even if a user-defined procedure is inlined to a program unit more than once, the size of data sections increases only once.

Even if a user-defined external or module procedure is inlined to all referenced points, the procedure is not deleted completely; outputinto the object file. This is because the procedure may be referenced from other program units. Therefore, internal procedures which areinlined to all referenced points are deleted completely; not output into object file.

If compile-time option -Koptmsg=2 is specified, messages are output listing the expanded places and procedures.

9.1.2.2 Optimization by modifying evaluation methodsWhen the compiler option -Kpreex is specified, the optimizations which move invariant expressions evaluated in advance are applied. Asa result, execution may abort because an instruction which should not be executed based on the logic of the program is executed.

However, this will not affect the precision of calculation results.

The effects of moving the execution location may appear upon referencing an intrinsic function or array element, and upon dividing.

The following example shows the effect of moving the execution location:

Example1: Using intrinsic functions DIMENSION A(100)...

DO I=1,100 IF(X.GT.0.0) THEN A(I)=ALOG(X) END IFENDDO...

->

DIMENSION A(100)...T=ALOG(X)DO I=1,100 IF(X.GT.0.0)THEN A(I)=T END IFENDDO...

[Description]

In the program on the left, there are no side effects because the function "alog" is called while the argument is greater than 0.0. In theprogram on the right, optimization (advance evaluation of invariant expressions) moves the function "alog(x)" out of the loop, and itis therefore always executed. Consequently, an error occurs when 'x' is less or equal than 0.0, and error message indicating that theargument of "alog" is incorrect is output.

Example2: Using array elements SUBROUTINE SUB(A,B,N)DIMENSION A(N),B(N)

DO I=1,N IF(J.LE.N) THEN A(I)=B(J)*F END IFENDDO...

->

SUBROUTINE SUB(A,B,N)DIMENSION A(N),B(N)

T=B(J)*FDO I=1,N IF(J.LE.N) THEN A(I)=T END IFENDDO...

[Description]

In the program on the left, the reference to array "b" will not exceed the boundary, because array element "b(j)" is referenced onlywhen "j" is less than "n". In the program on the right, however, optimization (advance evaluation of invariant expressions) has movedthe array "b(j)" out of the loop, and it is therefore always referenced regardless the value of "j". Consequently, "b(j)" may be outsideof the array boundary. If an array element outside the program boundary is referenced, the program will be aborted with an errormessage indicating that a memory protection exception (read) has occurred.

Example3: Division

- 223 -

SUBROUTINE SUB(A,N,F)DIMENSION A(N)

DO I=1,N IF(F.NE.0.0) THEN A(I)=B/F ELSE A(I)=0.0 END IFENDDO...

->

SUBROUTINE SUB(A,N,F)DIMENSION A(N)

T=B/FDO I=1,N IF(F.NE.0.0) THEN A(I)=T ELSE A(I)=0.0 END IFENDDO...

[Description]

In the program on the left, division is performed only when the variable "f" is not 0.0, so a division exception will not occur. In theprogram on the right, however, optimization (advance evaluation of invariant expressions) has moved the division "b/f" out of theloop, and therefore it is always executed regardless the value of "f". This means that the program may be aborted with a divisionexception error, if the division is performed while the value of "f" is 0.0.

Specifying -NRtrap option at the same time may cause exceptions that would not normally occur.

9.1.2.3 Advance evaluation of invariant expressionsIf the compiler option -Keval is specified, the optimization is performed that modifies how operations are evaluated for object programs,which can cause side effects in the calculation results. Operation exceptions, such as exponential overflow, may also occur. Changingdivision to multiplication is performed on division that uses floating point type variables.

The following example shows the effects of changing division to multiplication:

Example: Changing division-to-Multiplication operator DIMENSION A(100),B(100)...

DO I=1,100 A(I)=B(I)/CENDDO...

->

DIMENSION A(100),B(100)...

T=1.0/CDO I=1,100A(I)=B(I)*TENDDO...

[Description]

The variable "c" is not changed in the loop. With optimization (changing division to multiplication), the reciprocal of variable "c" iscalculated outside the loop, and the division within the loop is changed into a multiplication of the reciprocal. In other words, a singledivision operation is calculated from a division and a multiplication. If the result of "t=1.0/c" overflows the precision of the type ofvariable "c", the lower digit is rounded off, introducing a different result to the one obtained when optimization is not performed.Further, depending on the value of the variable "c", an exponential overflow exception may occur when calculating the reciprocaloutside the loop.

9.1.2.4 Loop StripingLoop striping superimposes an iteration of a loop on next several iterations of the loop. (The number of superimposed loop is called 'stripedlength'.) It can reduce overheads to iterate the loop, and it can promote software pipelining.

Example:

DO I=1,N A(I) = B(I) + C(I)ENDDO

If striped length is equal to 4, the -Kstriping=4 option is specified, instructions within a loop are expanded as shown in the following.

- 224 -

Example:

DO I=1,N,4 TMP_B1 = B(I) TMP_B2 = B(I+1) TMP_B3 = B(I+2) TMP_B4 = B(I+3) TMP_C1 = C(I) TMP_C2 = C(I+1) TMP_C3 = C(I+2) TMP_C4 = C(I+3) TMP_A1 = TMP_B1 + TMP_C1 TMP_A2 = TMP_B2 + TMP_C2 TMP_A3 = TMP_B3 + TMP_C3 TMP_A4 = TMP_B4 + TMP_C4 A(I) = TMP_A1 A(I+1) = TMP_A2 A(I+2) = TMP_A3 A(I+3) = TMP_A4ENDDO

Loop striping unrolls statements in a loop as with loop unrolling, so the sizes of object modules increase. It also increases compilationtime and memory requirement.

Note that performance may deteriorate by applying this optimization because the number of used registers increases.

9.1.2.5 XFILLXFILL instructions speed up following associated store instructions by securing the cache line.

It have similar characteristics to PREFETCH instructions in that they both secure the cache line for speed up, but differ in that the cacheline is used for "writing".

XFILL instructions are generated for stored array data which is in an innermost loop.

However, the compiler does not generate XFILL instructions for array data which is referred, accessed non-sequentially, or stored in IFconstruct.

When XFILL instructions are generated, prefetch instructions to the second level cache are not generated.

It is prohibited to apply the following optimizations because the loop is modified that data are always stored to the cache lines secured byXFILL instructions. It becomes performance can also decrease.

- Loop unrolling

- Loop striping

Performance may also be reduced under the following conditions:

- When there few iterations

- When the loop iteration count is smaller than the number of elements to be stored in the cache lines specified with -KXFILL=N.

However when the loop is modified as shown in Section 9.1.1.7 SIMD, even when XFILL optimization effect, message or optimizationinformation to which the above-mentioned optimizations are executed can be output.

The optimizations to generate XFILL instructions are suppressed by specifying compiler option -KNOXFILL.

Do not specify the -KXFILL option if performance is likely to decrease as a result.

It is usually desirable to control XFILL for each loop.

It is therefore recommended to specify the optimization control line "!OCL XFILL" rather than specifying the -K[NO]XFILL option whicheffects on the entire program.

9.1.2.6 UXSIMDUXSIMD(Unloop-eXamination SIMD) is optimization that not only loops but also widely applies SIMD extensions.

- 225 -

Do not depending on loop structure, UXSIMD function directs to choose operation and perform SIMD optimization.

The -Kuxsimd option with the -Ksimd and -KHPC_ACE options enables UXSIMD.

If -Knouxsimd option is set, UXSIMD optimization is not performed.

If -Koptmsg=2 option is specified, the message to show optimized sentence is output.

9.1.2.7 Loop Fission before and after IF constructs in loops(LFI)"LFI" is optimization that divides the loop into two or more loops when the loop can be divided before and after IF constructs in the loop.The performance might decrease by this optimization.

The improvement of the execution performance is expected in the following cases.

- IF constructs exists in the loop. AND,

- A lot of executable statements exist in the loop and IF constructs.

There is a possibility that the execution performance decreases in the following cases.

- IF constructs exists in the loop and there are few sentences in the loop. OR,

- IF constructs exists in the loop and there are a lot of IF constructs in the loop.

"LFI" perform the optimization when the -Kloop_fission and -Kloop_fission_if option together. If -Kloop_nofission is set, "LFT"optimization is not performed. If -Koptmsg=2 option is specified, the message to show optimized sentence is output.

An example of LFI is given below.

Example:LFI


REAL,DIMENSION(100) :: A,B,CINTEGER :: IDO I=1,100 A(I) = 1+A(I) ...

IF (COND > 50) THEN B(I) = B(I) / 2 ... END IF

C(I) = C(I)*2 ...END DO

->

REAL,DIMENSION(100) :: A,B,CINTEGER :: I1,I2,I3DO I1=1,100 A(I1) = 1+A(I1) ...END DODO I2=1,100 IF (COND > 50) THEN B(I2) = B(I2) / 2 ... END IFEND DODO I3=1,100 C(I3) = C(I3)*2 ...END DO

9.1.2.8 Loop VersioningThe loop versioning generates two loops to which optimization is applied and is not applied. The object program created by the compilerjudges the data dependency of array at execution time and selects either loop.

The optimization, such as SIMD, software pipelining, or automatic parallelization, is promoted by loop versioning.

The size of the object module and compile time may increase because the loop versioning generates two loops.

The execution performance may decrease because the processing for judgement is overhead.

The loop versioning is applied only to the innermost loop which contains a single array whose data dependency is unknown. The loopversioning may decrease the overhead because of the unknown dependency.

Moreover, the message to show where optimized by the loop versioning can be output by specifying the -Koptmsg=2 option.

The following example shows the loop versioning.

- 226 -

Example:

DO I=1,N A(I) = A(I+M) + B(I)ENDDO


IF ((1 .GT. N+M) .OR. (N .LT. 1+M)) THEN!OCL NORECURRENCE DO I=1,N ! Optimization A(I) = A(I+M) + B(I) ! ENDDO !ELSE DO I=1,N A(I) = A(I+M) + B(I) ENDDOENDIF

When the -Kloop_versioning option is specified, the compiler generates two loops with IF-constructs for judging the values of variable"N" and "M" so that the data dependency of the array "A" can be analyzed at execution time.

If the array "A" judges no data dependency, SIMD may be promoted for the loop.

9.1.2.9 CLONE OptimizationCLONE optimization generates conditional branches of variables for loops in order to promote other optimizations, such as full unrolling.CLONE optimization is applied by specifying the optimization control line. The calculation result is not guaranteed if the optimizationcontrol line CLONE is used incorrectly. See Section "9.10.4 Optimization Control Specifiers" and "9.18 Incorrectly Specified OptimizationIndicators" for more information about CLONE.

Example:

!OCL CLONE(N==10)DO I=1,N A(I) = IENDDO


IF (N==10) THEN DO I=1,10 A(I) = I ENDDOELSE DO I=1,N A(I) = I ENDDOENDIF

9.1.2.10 Link Time OptimizationThe following describes link time optimization.

9.1.2.10.1 Overview of Link Time Optimization

If the program consists of two or more source files, link time optimization is performed between the files at linking. Link time optimizationcan be controlled with the -Klto option. To apply link time optimization, -Klto option should be specified at both compilation and linking.Note that link time optimization is not applied when not frtpx command but ld command is used at linking.

- 227 -

Example:

- When the compilation and the linking are specified at a time:

$ frtpx -Kfast -Klto sample1.f90 sample2.f90 sample3.f90

- When the compilation and the linking are specified at different times:

$ frtpx -Kfast -Klto -c sample1.f90$ frtpx -Kfast -Klto -c sample2.f90$ frtpx -Kfast -Klto -c sample3.f90$ frtpx -Klto sample1.o sample2.o sample3.o

Link time optimization is performed for the object files which are compiled with the -Klto option. And, only files with an ".o" extensionare subject to link time optimization. Therefore, the library such as .so or .a is excluded from link time optimization.

The size of the compiled object files with the -Klto option is increased by the additional information for optimization. However, theadditional information for optimization is not included in the executable.

When the -Klto option is specified, the compiler outputs an intermediate file to the temporary directory by both of the compilation andthe linking. This file is deleted at the end of compilation. The default temporary directory is /tmp. The temporary directory can be changedby setting the environment variable TMPDIR.

Moreover, link time optimization can be applied to mixed language programming. But, link time optimization is effective only to thelanguage which the compilation command used at linking supports. See "Chapter 11 Mixed Language Programming" for the compilationcommand used at mixed language programming.

9.1.2.10.2 Notes on Link Time Optimization

The following notes apply when the link time optimization is used

- Link time optimization is a function which optimizes the program again before linking by using the information added to the objectfiles. Therefore, the memory or time which is needed to generate the executable may increase significantly.

- The following functions may be affected when link time optimization is applied.

- Information collection function at compilation

The optimization which is actually applied may be different from the following information at compilation:

- Compilation information which is output by the compiler option -Qp(-Nlst=p) or -Qt(-Nlst=t) (Optimization information)

- Optimization message which is displayed by the compiler option -Koptmsg.

See Section "4.1 Compilation Output" for compilation output of Fortran program.

- Program checks

The runtime check to the following Fortran program may not be executed:

- Validity of arguments and result of procedure checks (-Nquickdbg=argchk, -Ha)

- Subscript range checks (-Nquickdbg=subchk, -Hs)

- Undefined data reference checks (-Nquickdbg=undef, -Nquickdbg=undefnan, -Hu)

See Section "8.2.1 Debugging Check Functions" for debugging function of Fortran program.

- Profiler

The following information which the profiler outputs may not be guaranteed:

- Call graph information.

The function names in call graph information may include the function name which is generated internally by link timeoptimization.

- Source code information.

The each line cost may not be displayed correctly.

- 228 -

See the "Profiler User's Guide" for the profiler.

- Runtime information output function

The line number may not match the line number of the source program in information which runtime information output functionoutputs.

See the "Runtime Information Output Function" for runtime information.

- Trace back map

The line number of the statement where error occurred may not be correct.

See Section "4.2.2 Trace Back Map" for trace back map.

9.2 Notes of wrong erroneous programWhen optimization is performed on the following erroneous programs, execution results may differ with the optimization level. If resultsdiffer, check the following and correct the programs.

- Variables containing undefined values are referenced

- Variables containing values with unacceptable formats are referenced

- Array elements are referenced using index values that exceed array declarations

- There are violations of Fortran syntax

9.3 Effects of Environmental Change on OptimizationAs optimization functions operate based on the information from the source program that is converted in the compiler, differences in theversion of the compiler, the application of patches, and changes in the system environment may yield different results, even if the samesource program and compiler options are being used.

This may cause the following symptoms:

- Information in compile messages and compile log files regarding optimization and automatic parallelization show different compilationresults to previous ones.

- The results have some difference within the margin of error.

To know the difference of optimization results, compare compile messages or compile log file to before.

9.4 Effects of Changing the Execution OrderThe locations or numbers of the interruption on debugging may differ, because the execution order may be changed by removing a commonexpression from a loop or moving an invariant expression.

Note that the interruption caused by calculating constant values will occur only at runtime, because the calculations are not performed atcompilation.

9.5 Branching Target of Assigned GO TO StatementsThe label-lists of assigned GO TO statements must be accurate.

Assigned GO TO statements must not branch to a label that is not in the label-list.

Optimization assumes that each branch is directed to one of the labels in the label-list. Therefore, if there is branching to a label that is notin the list, unexpected results may occur.

If the label-list is omitted, the compiler analyzes the program and creates all possible statement labels as branch targets. Therefore, explicitlydefining branching labels will result in better optimization.

- 229 -

9.6 Effect of dummy argumentsNote that using incorrect syntax for dummy arguments will cause unexpected results.

Associate with an actual argument

Each dummy argument in a procedure is optimized as a separate variable. Therefore, updating a dummy variable may also update an actualargument and introduce an incorrect result. Even if the variables have different actual arguments, the result may be incorrect if the memoryis shared. According to the syntax, it is prohibited to redefine a value for a dummy argument both directly and indirectly in a subprogram.

The following example shows the effects of associating a dummy argument with an actual argument.

Example: Effect of connection with the same actual argument

[Description]

Dummy arguments "Y" and "Z" are optimized as different variables. Therefore, the optimization assumes that the value of "Z" in theDO loop is unchanged, and moves "Z+10.0" to outside the loop (moving an invariant expression). As a result, the results of the "Y=A/I" operation are not seen in the results of the "Z+10.0" operation. Further, even if the value of "Z" does change in the DO loop (Z+10.0is not moved), the results of the "Y=A/I" operation are not seen in the results of the Z+10.0 operation if the variables "Y" and "Z" areallocated to a register in the DO loop. This is because different registers are allocated to variable "Y" and variable "Z".

Linkage with data in a common block

Dummy arguments and variables in a common block are optimized under the assumption that they are not associated (that is, the user hasnot made grammar errors). For this reason, associating dummy arguments with variables in a common block will introduce an error.

The following example shows the effects of associating a dummy argument with data in a common block.

- 230 -

Example 1: Effect of linking data in a common block

[Description]

The dummy-data "X" and the variable "A" in the common block are optimized as independent variables. For this reason, theoptimization assumes that the "X*2.0" in the DO loop is a common expression, and replaces "Z=X*2.0" with Z=Y" (removing commonexpression). As a result, the value "Z" will not include the result of the expression "A=FLOAT(I)". The syntax prohibits assigning avalue by associating both a dummy argument and a common block with one instance at the same time. Only the dummy arguments orthe common block can be associated.

Example 2: Effect of associating a dummy argument with data in a common block

[Description]

The value of the instance "A" in the common block is changed by "SUB2" called by "SUB1". At the same time, dummy argument "X"associated with common block object "A" is also changed. However, the compiler cannot identify the changes and the expression "X+Y" before and after calling "SUB2" is subject to optimization (removing common expression). As a result, the "W=X+Y" after calling"SUB2" is replaced with "W=Z", and the "W" will no longer include the change in "X" made by "SUB2".

- 231 -

Association using an ENTRY statement

Suppose a subprogram that uses the ENTRY statement is called using an entry name, function name, or subroutine name, and then thedummy arguments and the corresponding actual arguments are associated. Referencing the variable may introduce a problem if thesubroutine is called using a different entry name, function name, or subroutine name, without instantiating the dummy argument as anactual argument. This is based on the assumption that the address of the actual argument is passed to the subprogram.

The following example shows the effect of the association using an entry name.

Example: Effect of association using a subentry

Explanation:

To be corrected, the call must be changed to

CALL ENT(X)

and the entry must be changed to

ENTRY ENT(X)

9.7 Restrictions on Inline ExpansionAs there are various restrictions, such as correspondence between actual and dummy variables, user-defined procedures selected for inlineexpansion are not always inline expanded. The procedures that can be inline expanded are external procedures, internal procedures, andmodule procedures.

The following describes the restrictions on inline expansion.

9.7.1 Restrictions on Called User-Defined Proceduresa. Excessive number of instructions

If the total number of instructions for a user-defined procedure exceeds the allowed maximum, the user-defined procedure is notinline expanded.The value of the allowed maximum can be changed with the parameter of the compile-time option -x. The default maximum valueis determined by the compiler.

b. If the size of an array exceeds the allowed limit

The user-defined procedure will not be inline expanded when the size of the array exceeds the maximum, excluding the commonblock of the user-defined procedure, dummy argument, and assigned variable (that is, the arrays that are allocated as static and localwithin the user-defined procedure).

- 232 -

The value of the allowed maximum can be changed using the parameter of the compile-time option -x. The default maximum valueis infinite.

c. Including a statement that restricts inline expansion

Procedures that include the following statements will not be inline expanded:

- ALLOCATE statement

- DEALLOCATE statement

- ASSIGN statement

- GO TO statement

- NAMELIST statement

- Variable group input-output statement

- SAVE statementWhen there is a SAVE statement that activates a variable, excluding the entirety of common blocks or elements of a commonblock (that is, variables that are local to the user-defined procedure), the user-defined procedure is not inline expanded.

- RETURN statementIf a user-defined procedure contains RETURN statement that returns an integer expression, it is not inline expanded.

- NULLIFY statement

- CONTAINS statement

d. When a procedure is subject to restrictions on inline expansion, the following procedures will not be inline expanded:

- Procedures with a VOLATILE attribute

- Functions that return an array, polymorphic, or polymorphic component

- Procedures whose result has a VOLATILE attribute

- Module procedures with a host module that has a PRIVATE variable.

- Module procedures with a host module that has an allocatable variable.

- Procedures using associated allocatable variable

- Module procedures defined by an ENTRY statement

- Procedures that has a procedure language binding specifier

- Procedures that have an ASYNCHRONOUS attribute

- Function result has a deferred length type parameter.

- Function result has a character type and the FUNCTION statement or ENTRY statement has a procedure language bindingspecifier.

e. When changing a value of a variable or an array that has an initial value.

- A user-defined procedure is not inline expanded if a variable of the user-defined procedure has an initial value defined by aDATA statement or type declaration statement, and the value may be changed directly or indirectly.

- A module procedure is not inline expanded if a variable or array has an initial value defined by a DATA statement or typedeclaration statement in its host module, and the value may be changed directly or indirectly.

f. An ENTRY statement is included

A subprogram that includes the following ENTRY statements is not inline expanded.

- The dummy parameters, defined with the function name or subroutine name, are different from ones defined on the entry name.

g. The -Knoalias=s option is specified

A procedure is not inline expanded if the procedure of either of the following conditions:

- 233 -

1. The calling and called procedures appear in different program units. And,the following either appears in the called procedure:

- Derived type variable with pointer attribute.

- Derived type variable contains a pointer component.

2. The calling and called procedures appear in different program units. And,the called procedure appears in module program unit. And,the following either appears in declaration part of the called module procedure:

- Derived type variable with pointer attribute.

- Derived type variable contains a pointer component.

3. The calling and called procedures appear in same program units. And,the dummy argument contained in the called procedure has a variable with a pointer attribute in the derived type variable.

9.7.2 Restrictions on the Relationship between Calling and Called ProgramUnits

a. When the actual and dummy arguments are mismatched in the following ways:

- The number of actual and dummy arguments differs.

- An actual argument and a dummy argument differ in attribute, type, or kind.

b. When the relationship between actual and dummy arguments is one of the following:

- A dummy argument is a dummy procedure name, and it corresponds to an actual argument which is an intrinsic function or aprocedure pointer

- A dummy argument is * and the actual argument it corresponds to is a statement label

- There is an array description (array expression) in the actual argument

- A dummy argument has the OPTIONAL attribute

- A dummy argument has the POINTER attribute

- A dummy argument has the ALLOCATABLE attribute

- A dummy argument has the VOLATILE attribute

- A dummy argument is an assumed-shape array

- A dummy argument has the VALUE attribute

- A dummy argument has the ASYNCHRONOUS attribute

- A dummy argument is polymorphic

- The component of derived type for a dummy argument is polymorphic

- A dummy argument is of parameterized derived type having assumed length type parameter

c. When there is an error in the way user-defined external procedures are referenced

- When an external function subprogram is referenced as an external subroutine subprogram

- When an external subroutine subprogram is referenced as an external function subprogram

- When an external function subprogram type is referenced with a different type

d. The user-defined external procedure and module procedure corresponds to the following criteria:

- A procedure specified as a local name by specifying a tentative name.

e. A reference to an elemental procedure.

- 234 -

9.7.3 Restrictions on User-Defined External Procedure NamesWhen there is a user-defined external procedure with the same name, a program error occurs and inline expansion is not applied:

- There is an external subroutine subprogram with the same name

- There is an external function subprogram with the same name

- There is a block data program unit with the same name

- There are two or more block data program units without names

When the compiler option -Koptmsg=2 is specified, a message is output indicating whether inline expansion is applied.

If inline expansion is not applied, a message is output indicating the reason.

9.7.4 Restrictions on the optionWhen -H option is specified, inline expansion is not applied.

9.8 Effects of Cray Pointer VariablesA pointer-based variable may be considered to share the memory of the variable with the address returned by the intrinsic function LOC.However, if the pointer contains the address of a common block object, a structure member, or a binding object, it is assumed that theaddress is shared with other data that belongs to the same common block object, structure, or binding object type.

In addition, assigning an address to a pointer-based variable without using the intrinsic function LOC may introduce an unexpected result.

Extra attention is required when a pointer-based variable is passed as a dummy argument, because the corresponding pointer is notconsidered to be sharing memory with the following variables:

- Other dummy arguments

- If other dummy arguments are pointer-based variables, the corresponding pointers

- Variables belonging to common blocks

9.9 Effects of Compiler Option -Kcommonpad[=N]If the compiler option -Kcommonpad is specified during a separate compilation, the compiler option -Kcommonpad must also be specifiedfor source files that include common blocks or using module with the same names. (Refer to example 1)

When compiling with the compiler option -Kcommonpad=N specified for multiple files, the value for N must be the same.

In addition, the program may not run properly if the compiler option -Kcommonpad is specified when the element of a common blockwith the same name is changed and used. (Refer to example 2)

If a source file which includes a common block or declaration part of module is compiled with the -Kcommonpad[=N] option, any othersource file which includes the same common block or using module must be compiled with the -Kcommonpad[=N] option, and value ofN must be identical when the source files are compiled separately. (Refer to example 1.)

If common blocks with the same name have different elements, the execution result is unpredictable when the -Kcommonpad[=N] optionis specified. (Refer to example 2.)

Example1:

a.f

COMMON/CMN/A,BINTEGER,DIMENSION(10) ::A=1,B=2...

b.f

COMMON/CMN/A,BINTEGER,DIMENSION(10) ::A,B

- 235 -

PRINT *,B...

Both a.f and b.f include the common block named cmn, so the compile-time option -Kcommonpad must be specified for both a.f andb.f.

Example2:

SUBROUTINE SUB_ACOMMON/CMN/A,BREAL(4),DIMENSION(10) ::A=1.0,B=2.0...END SUBROUTINE SUB_ASUBROUTINE SUB_BCOMMON /CMN/XCOMPLEX(4),DIMENSION(10) :: XPRINT *,X...END SUBROUTINE SUB_B

As the element of the common block cmn is different in subroutines sub_a and sub_b, operation will be incorrect if the compile-timeoption -Kcommonpad is specified.

9.10 Using Optimization control line (OCL)By providing relevant optimization information in the source program, the effectiveness of optimization can be increased. SpecifyingOCLs in the source program can improve optimizations.

9.10.1 Notation rules for optimization control linesOptimization control lines consist of lines beginning with the five characters "!OCL " (the fifth character is a space). Lines that begin with'!OCL' are usually treated as comment lines, but when the compiler option -Kocl is specified, they are used as optimization control linesand the operands become effective.

The notation rules are as follows:

!OCL i [,i] ....i: Optimization control specifierHowever, one or more blank spaces are necessary between "!OCL" and "i".

9.10.2 Description position of the optimization control lineThe position where the optimization control line can be inserted depends on the type of optimization indicator.

Optimization control lines are specified in program units: DO loop units, statement units, or array assignment statement units. Programunits, DO loop units, statement units, and array assignment statement units are defined as follows:

- Program unitThe top of each program unit.

- DO Loop unit

Immediately in front of DO statements, however, blank lines, comment lines, and other optimization control lines (that can be specifiedwithin DO-loop unit) can be specified between optimization control lines and the DO statement.

- Statement unitsImmediately in front of statements.

- Array assignment statement unitsImmediately in front of array assignment statements.

- 236 -

9.10.3 Wild Card SpecificationIn the operand of the following optimization control specifiers, a wild card may be specified for a variable name or a procedure name:

- NORECURRENCE

- NOVREC

See 12.2.3.2.5 Wild Card Specification for more information about Wild Card.

9.10.4 Optimization Control SpecifiersTable 9.1 lists optimization control specifiers that can be included in optimization control lines.

Table 9.1 List of optimization control specifiers that can be included in optimization control lines

Optimization control specifier Outline of meaning

Specifiable optimization controlline

P D S A

ARRAY_FUSION[,opt]

END_ARRAY_FUSION

Fusion of the array assignment statementwithin the range is promoted.

The specified optimization specifierdesignates to all the array assignmentstatement within the range.

- - - *

ARRAY_MERGE

(array1,array2[,array3]...) or

ARRAY_MERGE

(base_array:array1[,array2]...)

It specifies to merge the array or the arrayof few dimensions to the array of manydimensions.

base_array, array1, array2, ... are namesof arrays

* - - -

ARRAY_SUBSCRIPT

(array1[,array2]...)

It specifies to perform dimension shift ofarray.

array1, array2 ... are names of arrays.

* - - -

CACHE_SECTOR_SIZE

(l1_n1, l1_n2, l2_n1, l2_n2)

END_CACHE_SECTOR_SIZE

It specifies to direct the maximumnumber of ways of first level cache andsecond level cache in sectors 0 and 1 ofthe cache.

* - * -

CACHE_SECTOR_SIZE

(l2_n1, l2_n2)

END_CACHE_SECTOR_SIZE

It specifies to direct the maximumnumber of ways of second level cache insectors 0 and 1 of the cache.

* - * -

CACHE_SUBSECTOR_ASSIGN

(array1 [,array2...])

END_CACHE_SUBSECTOR

It directs to specify an array stored insector 1 of the cache.

* - * -

CLONE(var==n1[,n2]...)

It directs to generate conditionalbranches along to the arguments andgenerate loop copies and put them for theconditional blocks assuming that the varis invariant in the loops.

The conditional expressions are equals-to expressions which operands are varand n1[,n2]... of specified arguments.

Variable var is a name of integervariable.

- * - *

- 237 -



P D S A

Decimal from -9223372036854775808to 9223372036854775807 or integernamed constants can be specified asconstant value (n1[,n2]...).

EVALInstructs the system to performoptimization that changes the expressionevaluation sequence.

* * - *

NOEVAL Cancels the EVAL. * * - *

FISSION_POINT[(n)]It specifies to instruct to apply loopfission optimization.

- - * -

FLTLD Cancels the NOFLTLD. * * - *

NOFLTLDInstructs the system to performoptimization by using non-faulting loadinstruction which causes no signal.

* * - *

FP_CONTRACTInstructs the system to use M&Ainstruction.

* * - *

NOFP_CONTRACT Cancels the FP_CONTRACT. * * - *

FP_RELAXEDThe FP_RELAXED specifier instructs toapply optimizations which use reciprocalapproximation instructions.

* * - *

NOFP_RELAXED

The NOFP_RELAXED specifierinstructs to suppress optimizationswhich use reciprocal approximationinstructions.

* * - *

LOOP_BLOCKING(n)Designates blocking. Decimal from 2 to10000 can be specified as block size n.

* * - *

LOOP_NOBLOCKING Cancels the LOOP_BLOCKING. * * - *

LOOP_INTERCHANGE(var1,var2[,var3]...)

Designates that the sequence of a nestedDO loop is changed according to thespecified order.

var1, var2, var3... are inductionvariables.

- * - -

LOOP_NOINTERCHANGEThe LOOP_NOINTERCHANGEspecifier instructs to suppress loopinterchange optimizations.

- * - -

LOOP_NOFISSIONThe LOOP_NOFISSION suppress loopfission optimization.

* * - *

LOOP_NOFUSIONThis directs that fusion of the specifiedloop and the adjacent loops are notperformed.

* * - -

LOOP_PART_SIMDPerforms optimization which dividesloops and partially uses SIMDextensions.

* * - *

LOOP_NOPART_SIMDSuppresses optimization which dividesloops and partially uses SIMDextensions.

* * - *

- 238 -



P D S A

LOOP_VERSIONINGDesignates to perform optimization bythe loop versioning.

* * - -

LOOP_NOVERSIONING Cancels the LOOP_VERSIONING. * * - -

MFUNC[(level)]Designates to change the function to amulti-operation function. Decimal 1, 2 or3 can be specified as level.

* * - *

NOMFUNCDesignates canceling to change thefunction to a multi-operation function.

* * - *

NFOptimization using the Non-Faultingmode becomes effective.

* * - *

NONF Cancels the NF. * * - *

NOALIASDesignates that pointer variables andother variable do not share memory area.

* * - *

NOARRAYPAD(array1)

Designates canceling to execute paddingwith array element.

array1 is the name of array.

* - - -

NOVREC[(array1[,array2]...)]

Designates that arrays of recurrenceoperation do not exist in loop which canuse SIMD instructions. array1, array2, ...are names of arrays.

* * - *

NORECURRENCE[(array1[,array2]...)]

Designates that array section referencesdo not extend into loops of pluraliterations.

* * - *

PREEXInstructs the system to performoptimization by evaluating invariantfirst.

* * - *

NOPREEX Cancels the PREEX. * * - *

PREFETCHInstructs the system to perform theautomatic PREFETCH function of thecompiler.

* * - -

NOPREFETCH Cancels the PREFETCH. * * - -

PREFETCH_CACHE_LEVEL

(c-level)

Designates cache-level to prefetch ofdata. 1, 2 or all can be specified as c-level.

* * - -

PREFETCH_INFER

Instructs the system to be prefetch bydistance is guessed from existinginformation while be omitted of addresscalculation for prefetch.

* * - -

PREFETCH_NOINFER Cancels the PREFETCH_INFER. * * - -

PREFETCH_ITERATION(n)

Instructs the system to be target of aprefetch instruction would be data that isreferred to n iterations of a loop. Decimalfrom 1 to 10000 can be specified as n.

* * - -

PREFETCH_ITERATION_L2(n)Instructs the system to be target of aprefetch instruction would be data that isdefined or is referred to n iterations of a

* * - *

- 239 -



P D S A

loop. However, this target of prefetchinstruction while is prefetch only secondcache.

Decimal from 1 to 10000 can bespecified as n.

PREFETCH_READ

(name[,level={1|2}]

[,strong={0|1}])

Instructs the system to generate theprefetch instruction to the referred data.name is the name of array element. levelspecifies the level of cache to prefetch,and strong specifies to use strongprefetch or not.

- - * -

PREFETCH_WRITE

(name[,level={1|2}]

[,strong={0|1}])

Instructs the system to generate theprefetch instruction to the defined data.name is the name of array element. levelspecifies the level of cache to prefetch,and strong specifies to use strongprefetch or not.

- - * -

PREFETCH_SEQUENTIAL[(AUTO|SOFT)]

This directs that prefetch instructions aregenerated for array data that is accessedsequentially.

* * - *

PREFETCH_STRONGThis directs that the prefetch instructionfor the first level cache is to be the strongprefetch.

* * - *

PREFETCH_NOSTRONGThis directs that the prefetch instructionfor the first level cache will not be strongprefetch.

* * - *

PREFETCH_STRONG_L2This directs that the prefetch instructionsfor the second level cache are to be strongprefetch.

* * - *

PREFETCH_NOSTRONG_L2This directs that the prefetch instructionsfor the second level cache are not to bestrong prefetch.

* * - *

SHORTLOOP(n)The SHORTLOOP specifier instructs toapply some optimizations assuming itsiteration is few as the specified N.

* * - *

NOSHORTLOOPThe SHORTLOOP specifier instructs toapply some optimizations withoutassuming its iteration is few.

* * - *

SIMD[(ALIGNED |UNALIGNED)]

The SIMD specifier instructs to useSIMD instructions.

* * - *

NOSIMDThe NOSIMD specifier instructs not touse SIMD instructions.

* * - *

SIMD_LISTV

[(ALL | THEN | ELSE)]

This directs that the list vectorconversion is performed to the executionstatements in IF construct.

- - * -

SIMD_REDUNDANT_VL(n) This directs that SIMD with redundantexecutions for the SIMD width is applied

* * - *

- 240 -



P D S A

to loops whose number of iterations isequal to or less than n.

Decimal from 2 to 255 can be specifiedas n.

SIMD_NOREDUNDANT_VL This directs that SIMD with redundantexecutions for the SIMD width is notapplied.

* * - *

STRIPING[(n)]

This directs that the striping optimizationis performed. The striped length (numberof expansions) can be specified for n.

n indicates the stripe-length (number ofexpansions) which should be from 2 to100.

* * - *

NOSTRIPINGThis directs that the loop stripingoptimization is suppressed.

* * - *

SWP Designates software pipelining. * * - *

NOSWP Cancels the SWP. * * - *

UNROLL[(n)]

Designates unrolling a correspondingDO loop.

Decimal from 2 to 100 can be specifiedas the times of unrolling n.

- * - -

UNROLL('full') Designates full unrolling. - * - -

NOUNROLL Cancels the UNROLL. - * - -

UNSWITCHINGThe UNSWITCHING specifier instructsloop unswitching to IF construct.

- - * -

UXSIMD[(ALIGNED |UNALIGNED)]

The UXSIMD directs to use SIMDextensions by UXSIMD optimization.

* * - *

NOUXSIMD Cancels the UXSIMD. * * - *

XFILL[(n)]

It is directed to generate XFILLinstructions.

Decimal from 1 to 100 can be specifiedas the cache lines n.

- * - *

NOXFILLIt is directed not to generate XFILLinstructions.

- * - *

P: Program unit

D: DO loop unit

S: Statement unit

A: Array assignment statement unit

*: The optimization control specifier can be specified in the optimization control lines.

-: The optimization control specifier cannot be specified in the optimization control lines.

- 241 -

1. !OCL ARRAY_FUSION[,opt]!OCL END_ARRAY_FUSION

When ARRAY_FUSION specifier is put at the beginning of the target range and END ARRAY_FUSION specifier is put at the endof the range, array fusion is promoted in the range.

Not all array assignment statements within the range are fused even if the optimization control lines are used. Further, opt may bespecified in the optimization control specifier, and this will be effective in all array assignment statements within the range.The following shows an example:

Example:

REAL,DIMENSION(10) :: A,B,C ...!OCL ARRAY_FUSION * A=B+ ... *Useful range C=C+ ... *!OCL END_ARRAY_FUSION * ...

2. !OCL ARRAY_MERGE(array1,array2[,array3]...)

The ARRAY_MERGE specifier instructs to merge arrays. The arrays to be merged are explicit shape arrays.

See Section 9.15.2 The Restrictions of Merging Local Array Variables for information on restrictions.

This function will not be applied within specification expressions nor to the constant expressions of declaratives.

The following shows an example:

Example1:

!OCL ARRAY_MERGE(A,B,C) INTEGER,DIMENSION(101,102,103)::A,B,C CALL SUB(A) PRINT *,B(101,102,103) A=C A(101,102,103)=C(1,2,3)

When the -Kocl option is specified, the program above is considered as the following program.

INTEGER,DIMENSION(3,101,102,103)::TMP ! A, in 1 element of the 1st dimension, merge : TMP (1,101,102,103) ! B, in 2 element of the 1st dimension, merge : TMP (2,101,102,103) ! C, in 3 element of the 1st dimension, merge : TMP (3,101,102,103)CALL SUB(TMP(1,:,:,:)) ! An array name is changed to the array ! element which has shape. ...PRINT *,TMP(2,101,102,103)TMP(1,:,:,:) = TMP(3,:,:,:)TMP(1,101,102,103)=TMP(3,1,2,3)

Note:

TMP is a variable generated by the compiler.

Example2:

!OCL ARRAY_MERGE(A,B) PROGRAM MAIN INTEGER,DIMENSION(100,101,102) :: A,B ... CALL SUB(A) END SUBROUTINE SUB(C) INTEGER,DIMENSION(100) :: C

- 242 -

... END

When this feature is in effect, the arguments for the procedure SUB specified in MAIN may not be associated with the parameterC(100) in SUB, because a new element is appended in a dimension of A.

3. !OCL ARRAY_MERGE(base_array:array1[,array2]...)

This merges arrays to the base_array. This will affect intrinsic functions, for which the keyword arguments DIM and MASK canbe specified.

The arrays are merged into the final dimension of the base variable. The target arrays are explicit shape arrays with bounds that areconstant expressions (local arrays and common block objects).

See Section 9.15.2 The Restrictions of Merging Local Array Variables for information on restrictions (except for common-block-object items). This function will not be applied within specification expressions or the constant expressions of declaratives.


Example1:

!OCL ARRAY_MERGE(A:B,C) PROGRAM MAIN INTEGER,DIMENSION(10,11,12,2) :: A INTEGER,DIMENSION(10,11,12) :: B,C ... A=1 B=1 C(1,2,3)=1

The following shows how arrays are merged:

PROGRAM MAININTEGER,DIMENSION(10,11,12,4) :: AA=1A(:,:,:,3)=1A(1,2,3,4)=1

Example2:

!OCL ARRAY_MERGE(A:B,C) INTEGER,DIMENSION(1,2,3,4) :: A,X INTEGER,DIMENSION(1,2,3) :: B,C ... X=A ! An error occurs during compilation. A=RESHAPE((/(i,i=1,24)/),(/1,2,3,4/)) !The calculation result is not guaranteed because of expression ! shape difference from the right side to the left. PRINT *,UBOUND(A,4) ! 6 is output.

4. !OCL ARRAY_SUBSCRIPT(array1[,array2]...)

The ARRAY_SUBSCRIPT specifier instructs to perform dimension shift of arrays. The target array variables are allocatable arraysand explicit shape arrays with bounds that are constant expressions (local arrays, common block objects and dummy arguments).

See Section 9.14.2 Restrictions of the Target Variable of the Dimension Shift of the Array Declaration for information on restrictions(except the restriction for combination substance). This function will not be applied within specification expressions or the constantexpressions of declaratives.


Example1:

!OCL ARRAY_SUBSCRIPT(A) PROGRAM MAIN INTEGER,DIMENSION(101,102,103,5)::A CALL SUB(A)

- 243 -

... PRINT *,A(1,2,3,4)

If this type of program is compiled with the -Kocl option specified, the compiler will treat it as one of the following types ofprogram:

PROGRAM MAININTEGER,DIMENSION(5,101,102,103)::ACALL SUB(A)...PRINT *,A(4,1,2,3)

Example2:

!OCL ARRAY_SUBSCRIPT(A) PROGRAM MAIN INTEGER,DIMENSION(100,101,102,3) :: A ... CALL SUB(A) END SUBROUTINE SUB(B) INTEGER,DIMENSION(100) :: B ... END

In this example, the argument A(3,100,101,102) for SUB specified in MAIN will be merged with the parameter B(100) definedin SUB. However, the result may be incorrect because the dimension of the array is not identical.

Example3:

!OCL ARRAY_SUBSCRIPT(A) PROGRAM MAIN INTEGER,DIMENSION(1,2,3,4) :: A,B ... B=A ! An error occurs during compilation. A=RESHAPE((/(i,i=1,24)/),(/1,2,3,4/)) !The calculation result is not guaranteed because of expression shape ! difference from the right side to the left. PRINT *,UBOUND(A,4) !3 is output.

5. !OCL CACHE_SECTOR_SIZE(l1_n1, l1_n2, l2_n1, l2_n2)

!OCL END_CACHE_SECTOR_SIZE

!OCL CACHE_SECTOR_SIZE(l2_n1, l2_n2)

!OCL END_CACHE_SECTOR_SIZE

!OCL CACHE_SUBSECTOR_ASSIGN(array1 [,array2...])

!OCL END_CACHE_SUBSECTOR

See Section 9.17.2.1 Software control with optimization control rows for details.

6. !OCL CLONE(var==n1[,n2]...)

The CLONE specifier instructs to generate conditional branches along to the arguments and generate loop copies and put them forthe conditional blocks assuming that the var is invariant in the loops. The order of the generated branches is the order of thecorresponding arguments. This specifier promotes other optimizations, such as full unrolling. The execution result is not guaranteedin case that the value of variable var changes in the loop. For more details, refer to "9.18 Incorrectly Specified OptimizationIndicators".

The size of the object module and compile time may increase because this specifier makes copies of loops. This specifier is effectiveonly when the -O3 option is set.

Variable var is a name of integer variable whose kind value is 1, 2, 4, or 8.

- 244 -

You cannot specify following integer variable.

- Structure component

- Array, Array elements, or Substrings

- Pointer variables

- Allocatable variables

- Threadprivate variables

And, you cannot specify integer variable with following attributes.

- ALLOCATABLE

- CHANGEENTRY

- CONTIGUOUS

- DIMENSION

- EXTERNAL

- INTRINSIC

- PARAMETER

- POINTER

Decimal from -9223372036854775808 to 9223372036854775807 or integer named constants can be specified as constant values(n1[,n2]...). Do not specify any kind type parameter with an underscore '_' regardless of values.

You can use comma to enumerate constant values and colon to specify a range. The following two specifications are equivalent.

!OCL CLONE(var==1,3:5,7)

!OCL CLONE(var==1,3,4,5,7)

You can specify up to 20 values at once. When more than 20 values are specified, excess values are ignored. The followingspecification is limited in the number of values because it specifies 30 values.

!OCL CLONE(var==11:40)

The above specification is treated as the same as the following specification because excess values are ignored.

!OCL CLONE(var==11:30)

The following shows examples:

Example 1: SUBROUTINE SUB(A,N)...!OCL CLONE(N==10,20) DO I=1,N A(I) = I ENDDO...END SUBROUTINE

->

SUBROUTINE SUB(A,N)...IF (N==10) THEN DO I=1,10 A(I) = I ENDDO ELSE IF (N==20) THEN DO I=1,20 A(I) = I ENDDOELSE DO I=1,N A(I) = I ENDDOENDIF...END SUBROUTINE

- 245 -

Loop is copied with IF constructs in the order of the specification.

Example 2: SUBROUTINE SUB(A,N,M)...!OCL CLONE(N==10)!OCL CLONE(M==20) DO I=1,N A(I) = M ENDDO...END SUBROUTINE ->

SUBROUTINE SUB(A,N,M)...IF (N==10) THEN DO I=1,10 A(I) = M ENDDOELSE IF (M==20) THEN DO I=1,N A(I) = 20 ENDDOELSE DO I=1,N A(I) = M ENDDOENDIF...END SUBROUTINE

Loop is copied with IF constructs in the order of the multiple specifications.

Example 3: SUBROUTINE SUB(A,N,M)...!OCL CLONE(M==10)DO J=1,M!OCL CLONE(N==20) DO I=1,N A(I,J) = 0 ENDDOENDDO...END SUBROUTINE

->

SUBROUTINE SUB(A,N,M)...IF(M==10) THEN DO J=1,10 IF (N==20) THEN DO I=1,20 A(I,J) = 0 ENDDO ELSE DO I=1,N A(I,J) = 0 ENDDO ENDIF ENDDOELSE DO J=1,M IF (N==20) THEN DO I=1,20 A(I,J) = 0 ENDDO ELSE DO I=1,N A(I,J) = 0 ENDDO ENDIF ENDDOENDIF...END SUBROUTINE

CLONE specifiers for nested loops generate nested IF constructs.

7. !OCL EVAL

The EVAL specifier instructs to perform the optimization that changes the operation evaluation type. The following shows an example:

Example:

- 246 -

!OCL EVAL DO I=1,N A(I)=A(I)+B(I)+C(I)+D(I) ENDDO

->DO I=1,N A(I)=(A(I)+B(I))+(C(I)+ D(I))ENDDO

Optimization that changes the operation evaluation type in the target loop is performed.

8. !OCL NOEVAL

The NOEVAL specifier instructs to suppress the optimization that changes the operation evaluation type.


Example:

!OCL NOEVAL DO I=1,N A(I) = A(I) + B(I) + C(I) + D(I) ENDDO

Optimization that changes the operation evaluation type in the target loop is not performed.

9. !OCL FISSION_POINT[(n)]

The FISSION_POINT(n) specifier instructs to apply loop fission optimization.

The multiloop which is (n-1)th-parent from innermost loop is distributed.

- It is possible to specify "!OCL FISSION_POINT(n)" up to 30 lines in a loop, and these OCLs must have same n value.

An example is shown in the following.

Example: DO I=2,N DO J=2,N A(I,J) = A(I-1,J-1)!OCL FISSION_POINT(1) A(I,J) = A(I,J) + A(I-1,J) ENDDO ENDDO

->

DO I=2,N DO J=2,N A(I,J) = A(I-1,J-1) ENDDO DO J=2,N A(I,J) = A(I,J) + A(I-1,J) ENDDO ENDDO

10. !OCL FLTLD

The FLTLD specifier instructs that optimization using non-faulting load instructions (load where exceptions do not occur) is notperformed. This specifier is effective when the -KHPC_ACE option is set.

Example:

!OCL FLTLD DO I=1,N IF (M(I)) THEN A(I) = B(K) ENDIF ENDDO

Optimization using non-faulting load instructions within the target loop is not performed.

11. !OCL NOFLTLD

The NOFLTLD specifier instructs that optimization using non-faulting load instructions (load where exceptions do not occur) isperformed. This specifier is effective when the -KHPC_ACE option is set.

Example:

!OCL NOFLTLD DO I=1,N IF (M(I)) THEN

- 247 -

A(I) = B(K) ENDIF ENDDO

Optimization using non-faulting load instructions within the target loop is performed.

12. !OCL FP_CONTRACTThe FP_CONTRACT specifier instructs that Floating-Point Multiply-Add/Subtract instructions are to be used. Note that use ofthese instructions may introduce errors on the same level as rounding errors.

Example:

!OCL FP_CONTRACT DO I=1,N A(I) = A(I) + B(I) * C(I) ENDDO

13. !OCL NOFP_CONTRACTThe NOFP_CONTRACT specifier instructs that Floating-Point Multiply-Add/Subtract floating point operation instructions are notto be used.

Example:

!OCL NOFP_CONTRACT DO I=1,N A(I) = A(I) + B(I) * C(I) ENDDO

14. !OCL FP_RELAXED

The FP_RELAXED specifier instructs to perform high-speed reciprocal approximation operations for division or SQRT functions.This applies to single-precision and double-precision floating point calculations. This optimization may introduce rounding errors,compared to when the standard division or SQRT instructions are used.

Example:

!OCL FP_RELAXED DO I=1,N A(I) = SQRT(B(I) / C(I)) ENDDO

15. !OCL NOFP_RELAXEDThe NOFP_RELAXED specifier instructs to perform operations using division and SQRT without using reciprocal approximationinstructions, for single-precision and double-precision floating point calculations.An example is shown in the following.

Example:

!OCL NOFP_RELAXED DO I=1,N A(I) = SQRT(B(I) / C(I)) ENDDO

16. !OCL LOOP_BLOCKING(n)

The LOOP_BLOCKING specifier instructs to perform the blocking optimization. Specify the block size when blocking in n.


Example: !OCL LOOP_BLOCKING(80) DO J=1,N DO I=1,N A(I,J) = A(I,J) + B(J,I)

->

DO JJ=1,N,81 DO II=1,N,81 DO J=JJ,MIN(N,JJ+80) DO I=II,MIN(N,II+80) A(I,J) = A(I,J) + B(J,I) ENDDO

- 248 -

ENDDO ENDDO

ENDDO ENDDOENDDO

Loop blocking is performed with a block size of 80.

17. !OCL LOOP_NOBLOCKING

The LOOP_NOBLOCKING specifier instructs that loop blocking is suppressed.


Example:

!OCL LOOP_NOBLOCKING DO J=1,N DO I=1,N A(I, J) = A(I, J) + B(I, J) ENDDO ENDDO

18. !OCL LOOP_INTERCHANGE(var1,var2[,var3]...)

The LOOP_NOINTERCHANGE specifier instructs not to apply loop interchange. This allows DO loops to be interchanged in aspecified order. However, loop interchange is not performed when interchanging is not possible because it may introduce differentresult.

Example: !OCL LOOP_INTERCHANGE(I,J) DO I=1,M DO J=1,N A(I,J) = A(I,J) + B(J,I) ENDDO ENDDO

->

DO J=1,N DO I=1,M A(I,J) = A(I,J) + B(J,I) ENDDOENDDO

When LOOP_INTERCHANGE is not specified, parallelization occurs using DO-variable I, but with LOOP_INTERCHANGE,loops are interchanged and parallelization occurs using DO-variable J. When the variable N is sufficiently larger than the variableM, the parallelization becomes more effective.

19. !OCL LOOP_NOINTERCHANGE

The LOOP_NOINTERCHANGE specifier instructs that nested DO loops interchange is not performed.

Example:

!OCL LOOP_NOINTERCHANGE DO I=1,M DO J=1,N A(I,J) = A(I,J) + B(J,I) ENDDO ENDDO

20. !OCL LOOP_NOFISSION

The LOOP_NOFISSION specifier instructs to suppress loop fission optimization.


Example:

!OCL LOOP_NOFISSION DO I=2,N DO J=2,N A(I,J) = A(I-1,J-1) A(I,J) = A(I,J) + A(I-1,J)

- 249 -

ENDDO ENDDO

21. !OCL LOOP_NOFUSIONThe LOOP_NOFUSION specifier instructs that fusion of the specified loop and the adjacent loops are not performed. An example is shown in the following.

Example:

!OCL LOOP_NOFUSION DO I=1,N statement ENDDO DO J=1,N statement ENDDO DO K=1,N statement ENDDO

Fusion of the loop I and J are not performed. Fusion of the loop J and K are performed.

22. !OCL LOOP_PART_SIMD

The LOOP_PART_SIMD specifier instructs to perform optimization which divides loops and partially uses SIMD extensions.


Example:

!OCL LOOP_PART_SIMD DO I=2,N A(I) = A(I)-B(I)+LOG(C(I)) ! SIMD extensions can be applied to ! this statement D(I) = D(I-1)+A(I) ! SIMD extensions cannot be applied to ! this statement ENDDO


!OCL LOOP_PART_SIMD DO I=2,N A(I) = A(I)-B(I)+LOG(C(I)) ! SIMD execution ENDDO DO II=2,N D(II) = D(II-1)+A(II) ! no SIMD execution ENDDO

The loop is divided and SIMD extensions are applied to the part of the divided loop.

23. !OCL LOOP_NOPART_SIMD

The LOOP_NOPART_SIMD specifier instructs to suppress optimization which divides loops and partially uses SIMD extensions.


Example:

!OCL LOOP_NOPART_SIMD DO I=2,N A(I) = A(I)-B(I)+LOG(C(I)) D(I) = D(I-1)+A(I) ENDDO

Neither loop dividing nor using SIMD extensions is applied to the loop.

24. !OCL LOOP_VERSIONING

The LOOP_VERSIONING specifier instructs to perform optimization by the loop versioning.

- 250 -

The loop versioning is applied to only an innermost loop, and also the loop contains only one array with unknown data dependency.


Example:

DO I=1,N!OCL LOOP_VERSIONING DO J=1,N A(J) = A(J+M) + B(J,I) ENDDO ENDDO

The data dependency of array "A" is unknown at compile time because values of variable "N" and "M" are unknown at compiletime.

When LOOP_VERSIONING is specified, the compiler generates two loops with IF-constructs for judging the values of variable"N" and "M" so that the data dependency of the array "A" can be analyzed at execution time.

If the array "A" judges no data dependency, SIMD may be promoted for the loop.

25. !OCL LOOP_NOVERSIONING

The LOOP_NOVERSIONING specifier instructs that loop versioning is suppressed.

The LOOP_NOVERSIONING specifier is effective only for the innermost loop.


Example:

DO I=1,N!OCL LOOP_NOVERSIONING DO J=1,N A(J) = A(J+M) + B(J,I) ENDDO ENDDO

26. !OCL MFUNC[(level)]The MFUNC specifier instructs to perform optimization by converting intrinsic functions within DO loops, array assignmentstatement intrinsic functions, and operations within loops to multi-operation functions. Specify 1, 2, or 3 for level. Refer to thecompiler option -Kmfunc for more information on level. An example is shown in the following.

Example: When specified for loop

REAL(KIND=8),DIMENSION(10) :: A,B,C,D,E,F,G!OCL MFUNC DO I=1,10 A(I) = LOG(B(I)) E(I) = LOG(F(I)) ENDDO D = LOG(C) + LOG(G) ...

LOG(B(I)) and LOG(F(I)) are converted into multi-operation functions.

27. !OCL NOMFUNCThe NOMFUNC specifier instructs that optimization using multi-operation functions is not performed.An example is shown in the following.

Example:

REAL(KIND=8),DIMENSION(10) :: A,B,C,D,E,F,G!OCL NOMFUNC DO I=1,10 A(I) = LOG(B(I)) E(I) = LOG(F(I))

- 251 -

ENDDO D = LOG(C) + LOG(G) ...

LOG(B(I)) and LOG(F(I)) are not converted into multi-operation functions.

28. !OCL NF

The NF specifier instructs that optimization using the Non-Faulting mode is performed. This specifier is effective when the -KHPC_ACE2 option is set.An example is shown in the following.

Example:

!OCL NF DO I=1,N IF (M(I).NE.0) THEN A(I) = B(K) ENDIF ENDDO

29. !OCL NONF

The NONF specifier instructs that optimization using the Non-Faulting mode is not performed. This specifier is effective when the-KHPC_ACE2 option is set.An example is shown in the following.

Example:

!OCL NONF DO I=1,N IF (M(I).NE.0) THEN A(I) = B(K) ENDIF ENDDO

30. !OCL NOALIASThe NOALIAS specifier informs the compiler that no pointer variables share an address.The compiler can identify that the pointer variable will not share memory area with other pointers during the compilation, althoughthe content of the pointer is determined only at runtime. This promotes optimizations on the pointer variable. Note that this specifiermay not promote the optimization when the association status of the pointer variable is changed in the loop.An example is shown in the following.

Example:

REAL, DIMENSION(100), TARGET :: X REAL, DIMENSION(:), POINTER :: A,B A=>X(1:10) B=>X(11:20)!OCL NOALIAS DO I=1,10 B(I) = A(I) + 1.0 ENDDO END

Note:

If there are two or more pointer variables referring the same memory area in a loop with the NOALIAS specifier, the executionresult may be incorrect.

31. !OCL NOARRAYPAD(array1)The NOARRAYPAD specifier instructs that padding is not performed for the specified array elements. This is valid when thecompiler option

- 252 -

-Karraypad_const[=N] or -Karraypad_expr=N is set. An example is shown in the following.

Example:

!OCL NOARRAYPAD(A) PROGRAM MAIN INTEGER,DIMENSION(100,100) :: A COMMON /X/A PRINT *,UBOUND(A) PRINT *,SIZE(A,1),SIZE(A,2) CALL SUB END PROGRAM!OCL NOARRAYPAD(A) SUBROUTINE SUB INTEGER,DIMENSION(100,100) :: A COMMON /X/A PRINT *,UBOUND(A) PRINT *,SIZE(A,1),SIZE(A,2) END SUBROUTINE

Padding will not be performed to the array "A".

32. !OCL NOVREC[(array1[,array2]...)]

The NOVREC specifier is used to prescribe that arrays of recurrence operation do not exist in loop. The arrays in loop are usedSIMD instructions. But, the arrays in loop is not used SIMD instructions by kind of operation or loop structure.

The pointer array cannot be specified in array.

Example1:

REAL A(20),B(20)!OCL NOVREC DO I=2,10 A(I) = A(I+N) + 1 B(I) = B(I+M) + 2 ENDDO

The NOVREC specifier is used to prescribe that all arrays in loop is not recurrence operation. The operation of arrays A and Buses SIMD instructions.

Example2:

REAL A(20),B(20)!OCL NOVREC(A) DO I=2,10 A(I) = A(I+N) + 1 B(I) = B(I) + 2 ENDDO

The NOVREC specifier is used to prescribe that array A whose data dependency is unknown is not recurrence operation. Theoperation of arrays A and B uses SIMD instructions.

Refer to Section 9.18.2 Example of NOVREC specifier for invalid usages of NOVREC specifier

33. !OCL NORECURRENCE[(array1[,array2]...)]

The NORECURRENCE specifier instructs that each element of the arrays which are specified by this OCL is defined and referencedwithin the limits of iteration.

This enables some optimizations to be performed for the DO loop because the order of definitions and references become certain.

For example, the following optimizations are applied:

- Loop Slicing(Automatic Parallelization)

- SIMD

- 253 -

- Software Pipelining

Here, 'array' is an array, for which loop slicing can be performed. Wildcards can be specified in 'array'. When no array name isspecified, this specifier is applied to all of the arrays within the target range. See 12.2.3.2.5 Wild Card Specification for informationon specifying wildcards.

In "Example: Code without NORECURRENCE specifier", the compiler cannot determine whether loop slicing can be performedbecause the index of array-A is another array element L(I). For this reason, loop slicing is not performed for this DO loop.

Example: Code without NORECURRENCE specifier

REAL,DIMENSION(10000) :: A,BINTEGER,DIMENSION(10000) :: LDO I=1,10000 A(L(I)) = A(L(I)) + B(I)ENDDOEND

If it is certain that loop slicing will not affect array-A, specify NORECURRENCE as in "Example: Usage of NORECURRENCEspecifier" so that DO loop is sliced.

Example: Usage of NORECURRENCE specifier

REAL,DIMENSION(10000) :: A,B INTEGER,DIMENSION(10000) :: L !OCL NORECURRENCE(A) p DO I=1,10000 p A(L(I)) = A(L(I)) + B(I) p ENDDO END

Note:

Usage of this specifier is the programmer's responsibility.

When an element of the arrays which are specified by the OCL is defined or referenced in plural iterations, undesirableoptimizations may be applied. And the results will be unpredictable.

34. !OCL PREEXThe PREEX specifier instructs that optimization via advance evaluation of invariant expressions is performed. An example is shown in the following.

Example: !OCL PREEX DO I=1,N IF (M(I)) THEN A(I) = A(I) / B(K) ENDIF ENDDO

->

T = 1 / B(K) DO I=1,N IF (M(I)) THEN A(I) = A(I) * T ENDIF ENDDO

Advance evaluation of invariant expressions is performed.

35. !OCL NOPREEXThe NOPREEX specifier instructs to suppress optimizations which evaluate invariant expressions in advance. An example is shown in the following.

Example:

!OCL NOPREEX DO I=1,N IF (M(I)) THEN A(I) = A(I) / B(K) ENDIF ENDDO

Advance evaluation of invariant expressions is not performed.

- 254 -

36. !OCL PREFETCH!OCL NOPREFETCH!OCL PREFETCH_CACHE_LEVEL(c-level)!OCL PREFETCH_INFER!OCL PREFETCH_NOINFER!OCL PREFETCH_ITERATION(n)!OCL PREFETCH_ITERATION_L2(n)

This instructs to perform the automatic prefetch function of the compiler. The compiler automatically determines the mostappropriate position to insert and generates the prefetch instructions.

37. !OCL PREFETCH_READ (name[,level={1|2}] [,strong={0|1}])

The PREFETCH_READ specifier instructs to generate prefetch instructions for referenced data.

In name, specify the data (array element) referenced in the program. Expressions can also be specified in subscripts. level={1|2}directs which cache level is to have data prefetched. level=1 is default.

Specify whether to strong prefetch by specifying strong={0|1}. If strong=0 is specified, strong prefetch is not performed. If strong=1is specified, strong prefetch is performed. The strong=0 is default.

For data that is referenced and defined, use PREFETCH_WRITE.

38. !OCL PREFETCH_WRITE (name[,level={1|2}] [,strong={0|1}])

The PREFETCH_WRITE specifier instructs to generate prefetch instructions for defined data.

In name, specify the data (array element) defined in the program. Expressions can also be specified in subscripts. level={1|2} directswhich cache level is to have data prefetched. The level=1 is default.

Specify whether to strong prefetch by specifying strong={0|1}. If strong=0 is specified, strong prefetch is not performed. If strong=1is specified, strong prefetch is performed. strong=0 is default.


Example1: To generate prefetch instruction for A(INDX(I+16))

DO I = 1,N!OCL PREFETCH_READ(A(INDX(I+16))) T = T + A(INDX(I)) ENDDO

Example2: To generate prefetch instruction to the second level cache for A(INDX(I+16))

DO I = 1,N!OCL PREFETCH_READ(A(INDX(I+16)),level=2) T = T + A(INDX(I)) ENDDO

Moreover, when the automatic PREFETCH function of the compiler and it wants to control it exclusively because of the cacheuse efficiency, it is possible to achieve it by combining OCL as follows.

Example3: To generate prefetch instruction using the strong prefetch for A(INDX(I+16))

DO I = 1,N !OCL PREFETCH_READ(A(INDX(I+16)),strong=1) T = T + A(INDX(I)) ENDDO

Example4: To generate prefetch instruction for INDX(I+32) and A(INDX(I+16))

DO I = 1,N!OCL PREFETCH_READ(INDX(I+32))!OCL PREFETCH_READ(A(INDX(I+16))) T = T + A(INDX(I)) ENDDO

- 255 -

When controlling the use of the automatic prefetch provided by the compiler to improve cache utilization, use multiple OCLstatements as shown below.

Example5:

!OCL NOPREFETCH DO I = 1,N,4!OCL PREFETCH_READ(A(I+16))!OCL PREFETCH_READ(C(I+16))!OCL PREFETCH_WRITE(B(I+16)) B(I) = A(I) + C(I) B(I+1) = A(I+1) + C(I+1) B(I+2) = A(I+2) + C(I+2) B(I+3) = A(I+3) + C(I+3) ENDDO

39. !OCL PREFETCH_SEQUENTIAL [(AUTO|SOFT)]

The PREFETCH_SEQUENTIAL specifier instructs that prefetch instructions are generated for array data that is accessedsequentially. If neither AUTO nor SOFT is specified, AUTO is set by default.

1. !OCL PREFETCH_SEQUENTIAL(AUTO)

The PREFETCH_SEQUENTIAL(AUTO) instructs that the compiler automatically selects whether to use hardware-prefetchor to generate prefetch instructions for array data that is accessed sequentially in the loop.


Example:

!OCL PREFETCH_SEQUENTIAL (AUTO) DO I=1,N A(I) = B(I) ENDDO

Hardware prefetch is used for the target loop rather than generating prefetch instructions.

2. !OCL PREFETCH_SEQUENTIAL(SOFT)

The PREFETCH_SEQUENTIAL(SOFT) specifier instructs that the compiler does not use hardware-prefetch, but rathergenerates prefetch instructions for array data that is accessed sequentially in the loop.


Example:

!OCL PREFETCH_SEQUENTIAL (SOFT) DO I=1,N A(I) = B(I) ENDDO

Prefetch instructions are generated for the target loop rather than using hardware-prefetch.

40. !OCL PREFETCH_STRONG

The PREFETCH_STRONG specifier instructs that the prefetch instruction for the first level cache is to be the strong prefetch.


Example:

!OCL PREFETCH_STRONG DO I=1,N A(I) = B(I) ENDDO

Prefetch instruction for the first level cache generated for the target loop will be strong prefetch.

- 256 -

41. !OCL PREFETCH_NOSTRONG

The PREFETCH_NOSTRONG specifier instructs that the prefetch instruction generated for the first level cache will not be strongprefetch.


Example:

!OCL PREFETCH_NOSTRONG DO I=1,N A(I) = B(I) ENDDO

The prefetch instruction for the first level cache generated for the target loop will not be strong prefetch.

42. !OCL PREFETCH_STRONG_L2

The PREFETCH_STRONG_L2 specifier instructs that the prefetch instructions generated for the second level cache are to be strongprefetch.


Example:

!OCL PREFETCH_STRONG_L2 DO I=1,N A(I) = B(I) ENDDO

The prefetch instruction for the second level cache generated for the target loop will be strong prefetch.

43. !OCL PREFETCH_NOSTRONG_L2

The PREFETCH_NOSTRONG_L2 specifier instructs that the prefetch instructions generated for the second level cache are not tobe strong prefetch.


Example:

!OCL PREFETCH_NOSTRONG_L2 DO I=1,N A(I) = B(I) ENDDO

The prefetch instruction for the second level cache generated for the target loop will not be strong prefetch.

44. !OCL SHORTLOOP(n)

The SHORTLOOP specifier instructs to apply optimization for short loops assuming its iteration is few as the specified n. SeeSection -Kshortloop for more information about optimization for short loops. This specifier becomes effective for innermost loop.


Example:

SUBROUTINE SUB(A,B,C,N) REAL(KIND=8),DIMENSION(N) :: A,B,C ...!OCL SHORTLOOP(5) DO I=1,N A(I) = B(I) *C(I) ENDDO END SUBROUTINE

Optimization for short loops is performed assuming its iteration is 5 times.

45. !OCL NOSHORTLOOP

The SHORTLOOP specifier instructs to apply some optimizations without assuming its iteration is few.

- 257 -


Example:

SUBROUTINE SUB(A,B,C,N) REAL(KIND=8),DIMENSION(N) :: A,B,C ...!OCL NOSHORTLOOP DO I=1,N A(I) = B(I) *C(I) ENDDO END SUBROUTINE

Some optimizations are performed without assuming its iteration is few.

46. !OCL SIMD[(ALIGNED | UNALIGNED)]

The SIMD specifier instructs to perform SIMD optimization. Depending on the type of operation and the loop structure, however,SIMD optimization may not be performed.

ALIGNED and UNALIGNED are specifiers for controlling SIMD optimization to store floating point data. See Section 9.18.1Example of SIMD(ALIGNED) specifier for examples of common errors in specifying the SIMD(ALIGNED) specifier.


Example:

REAL(KIND=8),DIMENSION(10) :: A,B!OCL SIMD DO I=1,10 A(I) = A(I) + B(I) ENDDO

SIMD optimization is performed.

1. !OCL SIMD(ALIGNED)

This specifier is effective when the -KHPC_ACE option is set.

Fortran compilers perform distortion of loops to adjust the address boundaries in order to promote SIMD optimization. IfSIMD(ALIGNED) has been specified, loops are not modified to adjust address boundaries, and SIMD optimization isperformed for store instructions assuming that they are aligned to the following address boundaries:

- Single-precision floating point stores are aligned to 8-byte boundaries

- Double-precision floating point stores are aligned to 16-byte boundaries

See 5.2 Correct Data Boundaries for details about array boundaries the compiler allocates.

Example:

SUBROUTINE SUB(A,B,N) REAL(KIND=8),DIMENSION(N) :: A,B!OCL SIMD(ALIGNED) DO I=1,N A(I) = A(I) + B(I) ENDDO END SUBROUTINE

SIMD optimization is performed. SIMD optimization is also performed on store instructions.

2. !OCL SIMD(UNALIGNED)


When SIMD(UNALIGNED) has been specified, loops are not modified to adjust address boundaries, and the SIMDoptimization is not performed for store instructions with the assumption that store instructions are not aligned to 8 nor 16-byte boundaries.


- 258 -

Example2:

SUBROUTINE SUB(A,B,N) REAL(KIND=8),DIMENSION(N) :: A,B!OCL SIMD(UNALIGNED) DO I=1,N A(I) = A(I) + B(I) ENDDO END SUBROUTINE

SIMD optimization is performed. SIMD optimization is not performed on store instructions.

47. !OCL NOSIMD

The NOSIMD specifier instructs that SIMD optimization is not performed.

Example:

REAL(KIND=8),DIMENSION(10) :: A,B!OCL NOSIMD DO I=1,10 A(I) = A(I) + B(I) ENDDO

SIMD optimization is not performed.

48. !OCL SIMD_LISTV[(ALL | THEN | ELSE)]

The SIMD_LISTV specifier instructs to perform list vector conversion to the execution statements of the specified IF construct.ALL, THEN or ELSE is specifier for controlling the range of list vector conversion optimization. When the -O2 option or higheris set, and the -Ksimd=2 option or the same effect by the SIMD specifier is performed, this specifiers is effective.

See Section "9.1.1.7.3 List Vector Conversion" for the list vector conversion.

1. !OCL SIMD_LISTV(THEN)

The list vector conversion is performed to the statements in the subsequent block of the specified IF THEN statement.


Example:

!OCL SIMD DO I=1,N!OCL SIMD_LISTV(THEN) IF (A(I).EQ.0.0) THEN A(I) = A(I)+B(I) ELSE A(I) = A(I)+C(I) ENDIF ENDDO

2. !OCL SIMD_LISTV(ELSE)

The list vector conversion is performed to the statements in the subsequent block of the specified ELSE statement.


Example:

!OCL SIMD DO I=1,N!OCL SIMD_LISTV(ELSE) IF (A(I).EQ.0.0) THEN A(I) = A(I)+B(I) ELSE A(I) = A(I)+C(I) ENDIF ENDDO

- 259 -

3. !OCL SIMD_LISTV(ALL)

The list vector conversion is performed to the statements in the blocks of the specified IF THEN statement and ELSE statement.


Example:

!OCL SIMD DO I=1,N!OCL SIMD_LISTV(ALL) IF (A(I).EQ.0.0) THEN A(I) = A(I)+B(I) ELSE A(I) = A(I)+C(I) ENDIF ENDDO

49. !OCL SIMD_REDUNDANT_VL(n)

The SIMD_REDUNDANT_VL specifier instructs to validate SIMD with redundant executions for the SIMD width. This is appliedto loops whose number of iterations is equal to or less than n. When the number of iterations of the loops is not a multiple of theSIMD width, SIMD instructions with masks are used to generate loops which use SIMD extensions over the whole iterations.Decimal from 2 to 255 can be specified as n.

See Section "9.1.1.7.4 SIMD with Redundant Executions for the SIMD Width" for details.


Example:

!OCL SIMD_REDUNDANT_VL(7) DO I=1,M A(I) = B(I)+C(I) END DO

50. !OCL SIMD_NOREDUNDANT_VL

The SIMD_NOREDUNDANT_VL specifier instructs to invalidate SIMD with redundant executions for the SIMD width.

The following shows an example. This specifier is used to suppress the application to specific loops when theSIMD_REDUNDANT_VL specifier is specified for program units.

Example:

!OCL SIMD_REDUNDANT_VL(7)SUBROUTINE SUB(A,B,C) ...!OCL SIMD_NOREDUNDANT_VL DO I=1,M A(I)= B(I)+C(I) END DO ...END SUBROUTINE

51. !OCL STRIPING[(n)]

The STRIPING specifier instructs that the striping optimization is performed. The striped length (number of expansions) can bespecified for n.

A value between 2 and 100 can be specified for n. If a value was not specified for n, the compiler will automatically determine thevalue. When the iteration count of a loop in the source program is known, the expansion number determined by the compiler willbe used if a number that exceeds the iteration count is specified for n.

Example1:

!OCL STRIPING DO I=1,N

- 260 -

statement ENDDO

This specifies that loop striping is to be performed on the statement using the striping length determined by the compiler.

Example2:

!OCL STRIPING(4) DO I=1,N statement ENDDO

This specifies that loop striping is to be performed on the statement using the striping length 4.

Example3:

!OCL STRIPING(8) DO I=1,4 statement ENDDO

The specified stripe length (8) exceeds the loop iteration count (4), so the striped length automatically determined by the compilerwill be used.

Example4:

!OCL STRIPING(4) SUBROUTINE SUB ... ... DO I=1,N statement ENDDO DO I=1,N statement ENDDO

If this is specified at the start of the program unit, all the DO loops in the program are targeted.

52. !OCL NOSTRIPING

The NOSTRIPING specifier instructs that the loop striping optimization is suppressed.

Example:

!OCL NOSTRIPING DO I=1,N statement ENDDO

Loop striping is not performed.

53. !OCL SWP

The SWP specifier instructs that software pipelining is performed. This specifier (or indicator) is valid only when the -O2 optionor above is in effect at compilation.


Example:

!OCL SWP DO I=1,N statement ENDDO

54. !OCL NOSWP

The NOSWP instructs that software pipelining is suppressed.

- 261 -


Example:

!OCL NOSWP DO I=1,N statement ENDDO

55. !OCL UNROLL[(n)]

The UNROLL specifier instructs the number of unrolls for loop unrolling at n. This is described immediately before a DO constructwith a DO variable. n is unsigned number from 2 to 100. If a value was not specified for n, the compiler will automatically determinethe value.

Further, if the iteration count for a loop is known, the known number of iterations is given priority, even if a number is specified inn that exceeds the iteration count.

Note that only the DO-loop immediately after the optimization control line is subject to optimization control.

When the optimization control line has not been specified, the compiler will determine the loop unrolling expansion count basedon the iteration count, number and type of instructions in the loop, and the data type being used.


Example1:

SUBROUTINE SUB...!OCL UNROLL(8) DO I=1,N statement ENDDO...END SUBROUTINE

This specifies that the statement is to be unrolled 8 times.

Example2:

SUBROUTINE SUB...!OCL UNROLL(80) DO I=1,50 statement ENDDO...END SUBROUTINE

The specified expansion count (80) exceeds the loop iteration count (50), so the expansion count will be deemed to be 50.

56. !OCL UNROLL('full')

If 'full' is specified as the loop expansion count, instructions will be expanded as many times as the iteration count in the sourceprogram. Optimization is not performed if the iteration count is unknown.



Example1:

SUBROUTINE SUB...!OCL UNROLL('full') DO I=1,10 statement ENDDO

- 262 -

...END SUBROUTINE

This specifies that the statement is to be unrolled 10 times, the same as the loop iteration count.

Example2:

SUBROUTINE SUB...!OCL UNROLL('full') DO K=1,2!OCL UNROLL('full') DO J=1,2!OCL UNROLL('full') DO I=1,2 statement ENDDO ENDDO ENDDO...END SUBROUTINE

This specifies that the statement is to be unrolled 2*2*2 times.

57. !OCL NOUNROLL

The NOUNROLL specifier instructs that loop unrolling optimization is suppressed.



Example:

SUBROUTINE SUB...!OCL NOUNROLL DO I=1,N statement ENDDO...END SUBROUTINE

Loop unrolling is not performed.

58. !OCL UNSWITCHING

The UNSWITCHING specifier instructs loop unswitching to IF construct. This specifier must be described immediately before IFconstruct which is invariant in a loop. This specifier is not effective in the case of specified other than the above.


Example1:

SUBROUTINE SUB(A,B,C,X,N) REAL(KIND=8),DIMENSION(N) :: A,B,C INTEGER(KIND=8) :: X ... DO I=1,N!OCL UNSWITCHING IF (X == 0) THEN A(I) = B(I) ELSE A(I) = C(I) ENDIF ENDDO END SUBROUTINE

- 263 -

Loop unswitching to IF construct is performed.

Note that IF construct without this specifier is not the target of loop unswitching. The following shows an example:

Example2:

SUBROUTINE SUB(A,B,C,D,X,N) REAL(KIND=8),DIMENSION(N) :: A,B,C,D INTEGER(KIND=8) :: X ... DO I=1,N IF (X == 0) THEN A(I) = B(I)!OCL UNSWITCHING ELSE IF (X == 1) THEN A(I) = C(I) ELSE A(I) = D(I) ENDIF ENDDO END SUBROUTINE

Only IF construct with this specifier is the target of loop unswitching.

The memory consumption and the compilation time may increase drastically when loops to which loop unswitching is appliedinclude a lot of execution statements.

59. !OCL UXSIMD[(ALIGNED | UNALIGNED)]

The UXSIMD specifier instructs to SIMD optimization using UXSIMD optimization. However depending of kind of instruction,SIMD optimization is not performed.

ALIGNED and UNALIGNED are specifier for controlling SIMD optimization to store floating point data.

However, this does not direct to perform distortion of loops to adjust the address boundaries in order to promote SIMD optimization.

See "9.1.2.6 UXSIMD" for details about UXSIMD optimization.


Example:

!OCL UXSIMD DO I=1,N,2 A(I) = B(I) + C(I) A(I+1) = B(I+1) + C(I+1) ENDDO

SIMD optimization using UXSIMD is performed for instruction in the loop.

1. !OCL UXSIMD(ALIGNED)


If UXSIMD(ALIGNED) has been specified, SIMD optimization is performed for store instructions assuming that they arealigned to the following address boundaries:

- Single-precision floating point stores are aligned to 8-byte boundaries

- Double-precision floating point stores are aligned to 16-byte boundaries

Because SIMD optimization is assumed, the first element of a floating point array is aligned to the following address.

However arrays passed as arguments, pointer arrays, structure array and common block array have the possibility of notcorresponding to these.

- Single-precision floating point arrays are aligned to 8-byte boundaries

- Double-precision floating point arrays are aligned to 16-byte boundaries

- 264 -

Example:

!OCL UXSIMD(ALIGNED) DO I=1,N,2 A(I) = B(I) + C(I) A(I+1) = B(I+1) + C(I+1) ENDDO

SIMD optimization using UXSIMD is performed for instruction in the loop. SIMD optimization is also performed onstore instructions.

2. !OCL UXSIMD(UNALIGNED)


If UXSIMD(UNALIGNED) has been specified, SIMD optimization is not performed for floating point store instructions.

Example:

!OCL UXSIMD(UNALIGNED) DO I=1,N,2 A(I) = B(I) + C(I) A(I+1) = B(I+1) + C(I+1) ENDDO

SIMD optimization using UXSIMD is performed for instruction in the loop. SIMD optimization is not performed on storeinstructions.

60. !OCL NOUXSIMD

The UXSIMD directs not to use SIMD extensions by UXSIMD optimization.

Example:

!OCL NOUXSIMD DO I=1,N,2 A(I) = B(I) + C(I) A(I+1) = B(I+1) + C(I+1) ENDDO

UXSIMD optimization is not performed for instruction in the loop.

61. !OCL XFILL[(n)]

The XFILL specifier instructs to create an object to store data that will improve the cache utilization. Specify the number of linesto be placed on cache using n. A value between 1 and 100 can be specified for n. If a value was not specified for n, the compilerwill automatically determine the value.

For details about XFILL, see "9.1.2.5 XFILL".


Example1:

SUBROUTINE SUB(A,N)...!OCL XFILL DO I=1,N A(I) = 1 ENDDO...END SUBROUTINE

This specifies to improve the cache utilization for the data on four lines ahead.

Example2:

SUBROUTINE SUB(A,N)...

- 265 -

!OCL XFILL(1) DO I=1,N A(I) = 1 ENDDO...END SUBROUTINE

This specifies to improve the cache utilization for the data on one line ahead.

62. !OCL NOXFILL

The NOXFILL specifier instructs not to create an object to store data that will improve the cache utilization.


Example:

SUBROUTINE SUB(A,N)...!OCL NOXFILL DO I=1,N A(I) = 1 ENDDO...END SUBROUTINE

This specifies that optimization to improve cache utilization will not be performed.

9.11 Effects of Allocating on StackThe following shows the effects of allocating stack using the compiler option -Kautoobjstack, -Kauto, -Kthreadsafe, or -Ktemparraystack.

If the maximum stack area size is smaller than the memory required by the program, stack area may not be allocated, and the programmay terminate abnormally. In this case, increase the maximum stack area size.

To view and change the maximum, use the limit command on (csh) and the ulimit command on (sh). In addition, when executing as abatch request using Job Operation Software, the maximum size of the corresponding stack area of the batch queue (Per-process stack sizelimit) must be increased.

The following examples show the effects of specifying each of the options:

Example1: When "-Kautoobjstack" is specified

$ cat a.f90 SUBROUTINE SUB(N) REAL(KIND=8),DIMENSION(N) :: A A = 1. PRINT*, A(N) END SUBROUTINE

CALL SUB(2000000) END$ ulimit -s8192$ frtpx a.f90$ ./a.out 1.000000000000000$ frtpx a.f90 -Kautoobjstack$ ./a.out Segmentation Fault(core dumped)

Program terminated abnormally due to an insufficient stack area size.

Specify "unlimited" for the "stacksize" and compile again.

- 266 -

$ ulimit -s unlimited$ ulimit -sunlimited$ frtpx a.f90 -Kautoobjstack$ ./a.out 1.000000000000000

The program ended normally because the stack size was increased.

Example2: When "-Kauto" is specified

$ cat b.f90 REAL(KIND=8),DIMENSION(2000000) :: A A = 1. PRINT*, A(1) END$ ulimit -s8192$ frtpx b.f90 ; ./a.out 1.000000000000000$ frtpx b.f90 -Kauto ; ./a.out Segmentation Fault(core dumped)$ ulimit -s unlimited$ ulimit -sunlimited$ frtpx b.f90 -Kauto ; ./a.out 1.000000000000000

Example3: -Ktemparraystack" is specified

$ cat c.f90 SUBROUTINE SUB(A, N1, N2, M1, M2) REAL(KIND=8),DIMENSION(M2) :: A A(N1:N2) = A(M1:M2) + 1. PRINT*, A(1) END SUBROUTINE REAL(KIND=8),DIMENSION(2000000) :: A A = 0.

CALL SUB(A,1,1999999,2,2000000) END$ ulimit -s8192$ frtpx c.f90 ; ./a.out 1.000000000000000$ frtpx c.f90 -Ktemparraystack ; ./a.out Segmentation Fault(core dumped)$ ulimit -s unlimited$ ulimit -sunlimited$ frtpx c.f90 -Ktemparraystack ; ./a.out 1.000000000000000

9.12 Effects of Applying Optimization to Change Shape of ArrayIncreasing the maximum array size using the compiler option "-Karraypad_const[=N] " or "-Karraypad_expr=N" will affect the result ofthe functions UBOUND and SIZE. Further, if the "fdb" command or "SUBCHK" functionality is used, it will be as if the maximum sizehas been specified in the source program.

When using the compiler options "-Karraypad_const[=N]" and "-Karraypad_expr=N", the same options must be specified when compilingall the program units that make up the executable program.

- 267 -

This feature is applied to each individual array, without considering storage association and argument associations that are introduced bythe EQUIVALENCE statement and COMMON statement. For this reason, the correct result may not be obtained by using this functionso caution should be used.

Example1: -Karraypad_expr=N

PROGRAM MAINREAL,DIMENSION(1024*1024) :: ACALL SUB(A,1024,1024) :SUBROUTINE SUB(X,M,N)REAL,DIMENSION(M,N) :: X :

If this type of program is compiled with the "-Karraypad_expr=1" option, then it will be compiled as one of the following types ofprograms.

PROGRAM MAINREAL,DIMENSION(1024*1024+1) :: ACALL SUB(A,1024,1024) :SUBROUTINE SUB(X,M,N)REAL,DIMENSION(M+1,N) :: X :

If this feature is not used, A(1025) in MAIN is associated with X(1,2) in SUB. When it is used, A(1025) is associated with X(1025,1)and will derive a different result.

Example2: -Karraypad_const[=N]

INTEGER A(3,3),B(3,3)A(:,1)=B(1,:)A(:,2)=RESHAPE((/1,2,3/),(/3/))

If the "-Karraypad_const=1" option is specified on compile, the array shape on the right and left of the assignment will differ, and theresults may be incorrect.

9.13 Effects of generating prefetch on program performanceIf the -Kprefetch_sequential, -Kprefetch_stride, -Kprefetch_indirect, or -Kprefetch_conditional option is valid, it may decrease inperformance to generate prefetch instructions depending on the cache-efficiency of the loop, the presence or absence of branch instructions,and the complexity of subscripts.

9.14 The Dimension Shift of the Array DeclarationThis section explains dimension shift of array declarations.

9.14.1 Effects of the Dimension Shift of Array DeclarationThe dimension position of arrays is shifted by specifying the compiler option -Karray_subscript. This will affect intrinsic functions, forwhich the keyword arguments DIM and MASK can be specified.

When using the compiler options -Karray_subscript, the same options must be specified when compiling all the program units that makeup the executable program. The target array variables are allocatable arrays and explicit shape arrays with bounds that are constantexpressions (local arrays, common block objects, and dummy arguments). This function will not be applied within specification expressionsor the constant expressions of declaratives. This feature is applied to each individual array, without considering argument association. Forthis reason, the correct result may not be obtained by using this function so caution should be used.


- 268 -

Example1:

INTEGER,DIMENSION(101,102,103,5) :: ACALL SUB(A) :PRINT *,A(1,2,3,4)

If this type of program is compiled with the "-Karray_subscript" option, then it will be compiled as one of the following types ofprograms:

INTEGER,DIMENSION(5,101,102,103) :: ACALL SUB(A) :PRINT *,A(4,1,2,3)

Example2:

PROGRAM MAININTEGER,DIMENSION(100,101,102,3) :: A :CALL SUB(A)ENDSUBROUTINE SUB(B)INTEGER,DIMENSION(100) :: B :END

In this sample, the argument A(3,100,101,102) for SUB specified in MAIN will be merged with the parameter B(100) defined in SUB.However, the result may be incorrect because the dimensions of the arrays are not identical.

Example3:

INTEGER,DIMENSION(1,2,3,4) :: A : A=RESHAPE((/(i,i=1,24)/),(/1,2,3,4/))

! The right-hand shape of the substitution expression is different! from the left-hand one, so the execution result may not be correct. PRINT *,UBOUND(A,4)! 3 is output.

9.14.2 Restrictions of the Target Variable of the Dimension Shift of the ArrayDeclaration

Arrays that meet any of the following conditions cannot be the target of dimension shift of array declarations:

- Derived type, character type, or polymorphic

- Literal with name

- An array with a SAVE attribute (implicit SAVE attribute in declaration part of module is excluded) or TARGET attribute

- Pointer variable

- An array with an initial value

- Equivalence-object

- Namelist-group-object

- Automatic array

- An array with a PUBLIC attribute declared in the module specification part

- Address holding variable

- Function result

- 269 -

- Declaration in BLOCK

9.15 Merging Array VariablesThis section explains merging of array declarations.

9.15.1 Effects of Merging Array VariablesMultiple arrays are merged into one array by specifying the compiler option -Karray_merge.

The target array variables are explicit-shape arrays or automatic object arrays with bounds that are constant expressions, and automaticallyallocated arrays (Note). This function will not be applied within specification expressions or the constant expressions of declaratives. Thecompiler automatically adds the element of the first dimension and merges the elements. The merging order is undefined. This may causea compile error or may introduce an incorrect result, because the array names will be replaced with the merged array names. This featureis applied to each individual array, without considering lower bound of declaration and argument association. For this reason, using thisfunction may cause incorrect results.

(Note)

If the bound expression of automatic object array is the following condition, the automatic object array is not restricted.

1. Bound expression of declaration contain exactly one variable, or

2. Bound expression of declaration contain is constant expression.


Example1:

INTEGER,DIMENSION(101,102,103)::A,B,C CALL SUB(A) : PRINT *,B(101,102,103) A=C A(101,102,103)=C(1,2,3)

If this type of program is compiled with the -Karray_merge option specified, then it will be compiled as one of the following types ofprograms:

INTEGER,DIMENSION(3,101,102,103)::TMP ! A, in 1 element of the 1st dimension, merge : TMP (1,101,102,103) ! B, in 2 element of the 1st dimension, merge : TMP (2,101,102,103) ! C, in 3 element of the 1st dimension, merge : TMP (3,101,102,103) CALL SUB(TMP(1,:,:,:)) : PRINT *,TMP(2,101,102,103) TMP(1,:,:,:) = TMP(3,:,:,:) TMP(1,101,102,103)=TMP(3,1,2,3)

Note:


Example2:

PROGRAM MAIN INTEGER,DIMENSION(100,101,102) :: A : CALL SUB(A(1,1,1)) END SUBROUTINE SUB(C) INTEGER,DIMENSION(100) :: C : END

- 270 -

When this feature is used, the "A", which is an argument of SUB on MAIN, may not be associated with C(100), which is a parameterof SUB, because an element is appended in the first dimension of "A".

The compiler option-Karray_merge_common facilitates to merge the named common block object array.

The target array variables are explicit shape arrays within named common blocks with bounds that are constant expressions. In this case,a common block name is created. It is the following form:

_0_ <common block name>

All of the common-block-objects within the same named common block must be arrays with the same shape. This function will not beapplied within specification expressions or the constant expressions of declaratives. The compiler automatically adds the element of thefirst dimension and merges the elements. The merging order is undefined. This may cause a compile error or may introduce an incorrectresult, because the array name will be replaced with the merged array names. This feature is applied to each individual array, withoutconsidering argument association. For this reason, using this function may cause incorrect results.


Example1:

INTEGER,DIMENSION(101,102,103)::A,B,CCOMMON /COM/ A,B,CCALL SUB(A) :PRINT *,B(101,102,103)

If this type of program is compiled with the -Karray_merge_common option, then it will be compiled as one of the following types ofprograms:

INTEGER,DIMENSION(3,101,102,103)::TMPCOMMON /_0_COM/ TMPCALL SUB(TMP(1,:,:,:))PRINT *,TMP(2,101,102,103)

Note:


9.15.2 The Restrictions of Merging Local Array VariablesArrays that meet any of the following conditions cannot be the target of local array variable merge:

- Derived types or character types

- Literal with name

- An array with a SAVE attribute or TARGET attribute

- Pointer variable




- Declaration within the module

- Address holding variable

- Function result

- Common block objects

- Dummy arguments

- Specified in THREADPRIVATE statement

- Automatic array

- 271 -

- Declaration in BLOCK

Note:

If the bound expression of automatic object array is the following condition, the automatic object array is not restricted.

1. Bound expression of declaration contains exactly one variable, or

2. Bound expression of declaration is constant expression.

9.15.3 Restrictions of target variables on common block object array mergeArrays that meet any of the following conditions cannot be the target of common-block-object array merge:

- Derived types or character types

- An array with a TARGET attribute




- Automatic array

- An array within program units that include OpenMP statements

9.16 Multi-Operation Function

9.16.1 Calling of Multi-Operation FunctionMulti-operation functions are functions that improve performance by using intrinsic functions for calculations and operations on multiplearguments called at one time. By specifying compiler option -Kmfunc or optimization control line MFUNC, the compiler analyzes loopsand replaces them with multi-operation functions if possible.

If there are complex loops, and the compiler cannot analyze it, modify the program to call multi-operation function directly.

The following is required:

- The argument n which specifies the size must be an 8-byte integer.

- The external procedure names of the multi-operation functions are the external procedure names for the C language. For this reason,when called from a Fortran program, specifies to change how to process external procedure name by $pragma specifier.

- The argument check may be omitted in the multi-operation function for high-speed. Therefore, the user program may be terminatedabnormally when the special value (NaN and Inf, etc.) defined by IEEE 754 is input.

Table 9.2 lists the multi-operation functions that can be called directly from user programs.

Table 9.2 Multi-operation functions that can be called directly from user programs

function/operation type argument calling format content of calculation

ACOS real(4) v_acos(x,n,y) y(i) = acos(x(i))

real(8) v_dacos(x,n,y) y(i) = dacos(x(i))

ASIN real(4) v_asin(x,n,y) y(i) = asin(x(i))

real(8) v_dasin(x,n,y) y(i) = dasin(x(i))

ATAN real(4) v_atan(x,n,y) y(i) = atan(x(i))

real(8) v_datan(x,n,y) y(i) = datan(x(i))

ATAN2 real(4) v_atan2(x1,x2,n,y) y(i) = atan2(x1(i),x2(i))

real(8) v_datan2(x1,x2,n,y) y(i)=datan2(x1(i),x2(i))

ERF real(4) v_erf(x,n,y) y(i) = erf(x(i))

- 272 -

function/operation type argument calling format content of calculation

real(8) v_derf(x,n,y) y(i) = derf(x(i))

ERFC real(4) v_erfc(x,n,y) y(i) = erfc(x(i))

real(8) v_derfc(x,n,y) y(i) = derfc(x(i))

EXP real(4) v_exp(x,n,y) y(i) = exp(x(i))

real(8) v_dexp(x,n,y) y(i) = dexp(x(i))

complex(8) v_cdexp(x,n,y) y(i) = cdexp(x(i))

EXP10 real(4) v_exp10(x,n,y) y(i) = exp10(x(i))

real(8) v_dexp10(x,n,y) y(i) = dexp10(x(i))

LOG real(4) v_alog(x,n,y) y(i) = alog(x(i))

real(8) v_dlog(x,n,y) y(i) = dlog(x(i))

LOG10 real(4) v_log10(x,n,y) y(i) = log10(x(i))

real(8) v_dlog10(x,n,y) y(i) = dlog10(x(i))

SIN real(4) v_sin(x,n,y) y(i) = sin(x(i))

real(8) v_dsin(x,n,y) y(i) = dsin(x(i))

COS real(4) v_cos(x,n,y) y(i) = cos(x(i))

real(8) v_dcos(x,n,y) y(i) = dcos(x(i))

SIN & COS real(4) v_scn(x,n,y,z) y(i) = sin(x(i))

z(i) = cos(x(i))

real(8) v_dscn(x,n,y,z) y(i) = dsin(x(i))

z(i) = dcos(x(i))

exponentiation real(4) v_arxr(x1,x2,n,y) y(i) = x1(i)**x2(i)

real(4) v_arxr1(x,a,n,y) y(i) = x(i)**a

real(8) v_adxd(x1,x2,n,y) y(i) = x1(i)**x2(i)

real(8) v_adxd1(x,a,n,y) y(i) = x(i)**a

Note:

When calling multi-operation functions directly from user programs, the memory used for the argument to return the results of theoperations must be separate from the memory used for the other arguments. If the areas overlap, the result may be incorrect.

The following shows the example of calling multi-operation functions.

Example1:

INTEGER,PARAMETER :: N=1000REAL(KIND=8) :: AREAL(KIND=8),DIMENSION(N) :: X,Y,Z :A = 0.2D0DO I=1,N Y(I) = DEXP(X(I)) Z(I) = X(I)**AENDDO

Invocation of multi-operation function

INTEGER(KIND=8),PARAMETER :: N=1000REAL(KIND=8) :: AREAL(KIND=8),DIMENSION(N) :: X,Y,ZEXTERNAL V_DEXP !$PRAGMA C(V_DEXP)

- 273 -

EXTERNAL V_ADXD1!$PRAGMA C(V_ADXD1) :A = 0.2D0CALL V_DEXP(X,N,Y)CALL V_ADXD1(X,A,N,Z)

Example2:

If a function is called within an IF construct, multi-operation functions can be used by storing only the elements that need calculationin arrays.

INTEGER,PARAMETER :: N=1000REAL(KIND=8) :: X,Y,ZREAL(KIND=8),EXTERNAL :: FUNC :DO I=1,N X = .. IF (X.GT.A) THEN Y = Y + SIN(X) Z = Z + COS(X) END IFENDDO

Invocation of multi-operation function

INTEGER(KIND=8),PARAMETER :: N=1000INTEGER(KIND=8) :: JREAL(KIND=8) :: X,Y,ZREAL(KIND=8),DIMENSION(N) :: WX,WY,WZEXTERNAL V_DSCN !$PRAGMA C(V_DSCN) :J = 0DO I=1,N X = .. IF (X.GT.A) THEN J = J + 1 WX(J) = X END IFENDDOCALL V_DSCN(WX,J,WY,WZ)DO I=1,J Y = Y + WY(I) Z = Z + WZ(I)ENDDO

9.16.2 Effects of Compiler Option -Kmfunc=3If the compile option -Kmfunc=3 is specified, the program may terminate abnormally as a side effect of changing loops that include IFconstructs into multi-operation functions. The following shows an example of the effects of -Kmfunc=3:

Example: Loops that include IF constructs

SUBROUTINE SUB(A,B) DIMENSION A(1000),B(1000) . . DO I=1,2000 IF (I.LE.1000) THEN A(I) = COS(B(I)) ENDIF ENDDO

[Description]

- 274 -

In the above example, the value of the subscript of array B never exceeds the declared range because the intrinsic function COS, ofwhich the actual argument is the array element, is called if the condition I.LE.1000 holds.

By applying the multi-operation function optimization, the DO loop references an array element for each iteration of the loop, anduses it as an argument of multi-operation function regardless the result of IF-construct. Access may therefore exceed the array bounds.This will cause an abnormal termination of the program, with a message informing of a memory protection exception (READ).

To prevent this, do not specify the compiler option -Kmfunc=3.

9.17 Software Control of Sector Cache

9.17.1 Using Sector CacheSector cache is a mechanism to prevent reusable data on the cache from being driven out by non-reusable data. In particular, sector cacheis effective for parallel execution when the shared L2 cache is accessed by multiple CPUs. By using sector cache, reusable data is separatedfrom non-reusable data on the cache. There are two methods to control sector cache with software; environment variables and optimizationcontrol lines. Both are methods of specifying the maximum number of ways for each sector, which are detailed below.

Performance will not be improved if the maximum number of ways is specified without consideration for the size of the array that is tobe protected from being driven out. The cache utilization will decline, and may cause a reduction in speed. Further, even if software controlof the sector cache is not used, cache control with LRU (Least Recently Used) still works. Thus speed may not improve even if this isspecified. One effective way to improve the performance of sector caches is to determine the reusable array size and specify the numberof ways for sector 1 to store the array, after ensuring L2 cache miss has occurred by confirming the PA information. Note that in order touse sector caches, the compile-time option -Kdalign must be set.

9.17.2 How to Control Sector Cache by SoftwareThere are two methods to control sector caches with software in this processor, as follows:

- Optimization control rows

- Environment variables and optimization control rows

9.17.2.1 Software control with optimization control rowsThere are three optimization control specifiers used for software control of sector caches, as follows:

- CACHE_SECTOR_SIZE(l2_n1, l2_n2) to END_CACHE_SECTOR_SIZE

- CACHE_SECTOR_SIZE(l1_n1, l1_n2, l2_n1, l2_n2) to END_CACHE_SECTOR_SIZE

- CACHE_SUBSECTOR_ASSIGN(array1 [,array2...]) to END_CACHE_SUBSECTOR

The CACHE_SECTOR_SIZE specifier directs the maximum number of ways in sectors 0 and 1 of the cache.

When 4 arguments are specified, they direct the number of all ways for caches: l1_n1, for sector 0; l1_n2, for sector1; l2_n1, for sector0of the L2 cache; l2_n2, for sector1 of the L2 cache.

When two arguments are specified, they direct that the maximum number of ways for sector 0 in the L2 cache it will be l2_n1, and forsector 1 of the L2 cache it will be l2_n2.

To direct in function, the CACHE_SECTOR_SIZE specifier must be used in procedure line. In this case, the CACHE_SECTOR_SIZEspecifier is valid in the scope of the function.

To direct a part of function, the range must be indicated by the CACHE_SECTOR_SIZE and END_CACHE_SECTOR_SIZE specifiersin statement lines. Note that the range must be not nested.

However, combined use of procedure line and statement line is accepted.

The arguments l1_n1, l1_n2, l2_n1, and l2_n2 must be as follows;

- 0 ≤ l1_n1 ≤ the maximum number of ways in the L1 cache



- 275 -


In order to ensure that the data in sector 0 does not drive out the data in sector 1, it is recommended to configure {l1_n1+ l1_n2 = maximumnumber of ways in L1 cache}, and {l2_n1+ l2_n2 = maximum number of ways in L2 cache}.

The CACHE_SUBSECTOR_ASSIGN specifier directs to specify an array stored in sector 1 of the cache.

However, it becomes effective only when the array of numeric type or logical type is specified.

To direct in function, the CACHE_SUBSECTOR_ASSIGN specifier must be used in procedure line. In this case, theCACHE_SUBSECTOR_ASSIGN specifier is valid in the scope of the function.

To direct a part of function, the range must be indicated by the CACHE_SUBSECTOR_ASSIGN andEND_CACHE_SUBSECTOR_ASSIGN specifiers in statement lines.

Note that the range must be not nested.

However, combined use of procedure line and statement line is accepted.


Example: Control sector caches by OCLs

SUBROUTINE OCL_SECTOR(A,B,N,N2) REAL(KIND=8),DIMENSION(N,N) :: A REAL(KIND=8),DIMENSION(N,N,N2) :: B! Specify the maximum number of ways for each sector in the level 2 cache! Specify the maximum number of ways in sector 0 = 2, ! the maximum number of ways in sector 1 = L2_MAX-2! L2_MAX = maximum number of ways in the L2 cache!OCL CACHE_SECTOR_SIZE (the maximum number of ways in sector 0, the maximum number of ways in sector 1)! Optimization control specifier to specify an array stored in sector 1! Array A is specified for Sector 1!OCL CACHE_SUBSECTOR_ASSIGN (A) DO K=1,N2 DO J=1,N DO I=1,N A(I,J)=A(I,J)+B(I,J,K) ENDDO ENDDO ENDDO! End scope of CACHE_SUBSECTOR_ASSIGN.!OCL END_CACHE_SUBSECTOR! End scope of CACHE_SECTOR_SIZE.! Return the maximum number of ways for each sector to the value ! they were before specifying CACHE_SECTOR_SIZE.!OCL END_CACHE_SECTOR_SIZE END SUBROUTINE

9.17.2.2 Software control with environment variables and optimization control rowsThe initial values for the maximum number of ways for each sector can also be specified with the following environment variable. Themaximum number of ways for the object program at startup can be specified with environment variables.

FLIB_SCCR_CNTL

This environment variable specifies to use sector cache. The values that can be set in the environment variable are shown below, alongwith their meanings. TRUE (use sector cache) is default.

- TRUE: use sector cache.

- FALSE: do not use sector cache.

FLIB_L2_SECTOR_NWAYS_INIT

This environment variable specifies the initial maximum number of ways for sector caches. The values that can be set in the environmentvariable are of the same format as that used in the CACHE_SECTOR_SIZE optimization specifier; n1 and n2. This is valid only when

- 276 -

FLIB_SCCR_CNTL is TRUE. 0,0 is default. The values that can be specified are of the same format as the CACHE_SECTOR_SIZEoptimization specifier, n1, n2, and in range 0 ≤n1, n2 ≤{maximum number of ways} respectively. In order to ensure that the data insector 0 does not drive out the data in sector 1, it is recommended to specify n1+n2= maximum number of ways.

The following example specifies 2,10 as the value of environment variable FLIB_L2_SECTOR_NWAYS_INIT before starting theprogram, which enables the software control to store Array A on the sector 1 with maximum number of ways for sector 1 to be 10.

Example: Ways to control sector caches with software (2)

SUBROUTINE OCL_SECTOR(A,B,N,N2) REAL(KIND=8),DIMENSION(N,N) :: A REAL(KIND=8),DIMENSION(N,N,N2) :: B! Optimization control specifier to specify an array stored in sector 1! Array A is specified for Sector 1!OCL CACHE_SUBSECTOR_ASSIGN (A) DO K=1,N2 DO J=1,N DO I=1,N A(I,J)=A(I,J)+B(I,J,K) ENDDO ENDDO ENDDO! End scope of CACHE_SUBSECTOR_ASSIGN.!OCL END_CACHE_SUBSECTOR END SUBROUTINE

9.17.2.3 Behavior when Invalid Value is specifiedIf a value that exceeds the maximum number of ways is specified for n1 or n2 in the optimization indicator CACHE_SECTOR_SIZE andthe environment variable FLIB_L2_SECTOR_NWAYS_INIT, the results will be the same as if the actual maximum number of ways isspecified. If a number less than 0 is specified, that optimization control line is ignored. If the specified value is outside the range of two-byte signed integer, operation will be incorrect.

9.18 Incorrectly Specified Optimization IndicatorsWhen optimization control lines are incorrectly specified, the execution result may not be incorrect.

9.18.1 Example of SIMD(ALIGNED) specifierSIMD optimization can be used when double-precision floating point stores are on 16-byte boundaries. However, if SIMD(ALIGNED)specifier is specified for double-precision floating point stores that are not on 16-byte boundaries, the program will terminate abnormallywith SIGSEGV.

See 5.2 Correct Data Boundaries for details about array boundaries the compiler allocates.


Example1:

REAL(KIND=8),DIMENSION(10) :: A COMMON //N,A!OCL SIMD(ALIGNED) DO I=1,10 A(I) = ... ENDDO

Array A may not be on 16-byte boundary because that A is not first element of a COMMON block.

Example2:

REAL(KIND=8),DIMENSION(10) :: A COMMON //A!OCL SIMD(ALIGNED) DO I=2,10

- 277 -

A(I) = ... ENDDO

While the array A is on 16-byte boundary, the element of array referenced first in this loop is the second element of array A, and thatelement is not on 16-byte boundary.

9.18.2 Example of NOVREC specifierWhen array A is not recurrence data and NOVREC specifier is effective, compiler uses SIMD instructions. When array A is recurrencedata and NOVREC specifier is effective, compiler uses SIMD instructions. But the execution result is unpredictable.

Example:

!OCL NOVREC DO I=1,100 A(I)=A(I+M)+... ENDDO

9.18.3 Example of CLONE specifierIn the copied loops under the generated branches, the specified variables are treated as invariants in the loops. Thus, the execution resultis not guaranteed in case that the value of the variables changes in the loop.

Example 1:

INTEGER,DIMENSION(32)::AINTEGER,TARGET::NINTEGER,POINTER::MM=>N!OCL CLONE(N==10)DO I=1,32 M=5 A(I) = NENDDO

Do not specify a variable which is changed in the loop because the execution result is not guaranteed.

Example 2:

INTEGER::N!OCL CLONE(N==10)DO I=1,32 CALL SUB(N)ENDDO

Do not specify a variable which may be changed in the loop because the execution result is not guaranteed.

9.19 Attentions for Mixed Language ProgrammingThe following describes the notes when programming with different languages in this processor.

9.19.1 Address parameters received in the C languageIf an external C program stores the addresses of the parameters, the result may be incorrect because such operation is unsupported and theoptimization is performed without considering it. (Refer to example 1)

In this case, to obtain the correct result, pass the address explicitly using arguments rather than saving the parameter address in the function.(Refer to example 2)

Alternatively, pass the parameter address by specifying "%VAL(LOC(argument))" as the argument of the function caller in the Fortranprogram in Example 1. (Refer to Example 3)

- 278 -

Example1: Incorrect saving of pointer

int fun_(int *flag, int *p){ static int *save; if (*flag == 0) { save = p; /* Saving address dummy argument. */ return 0; } else { return *save; }}

INTEGER(KIND=4)::I,J INTEGER(KIND=4),EXTERNAL::FUN J = 0 I = FUN(0, J) J = 1 I = FUN(1, 0) PRINT *, I, J END

Example2: Format for not saving pointer

int fun_(int *flag, int *p){ if (*flag == 0) { return 0; } else { return *p; }}

INTEGER(KIND=4)::I,J INTEGER(KIND=4),EXTERNAL::FUN J = 0 I = FUN(0, J) J = 1 I = FUN(1, J) PRINT *, I, J END

Example3: Format for explicit passing of address

INTEGER(KIND=4)::I,J INTEGER(KIND=4),EXTERNAL::FUN J = 0 I = FUN(0, %VAL(LOC(J))) J = 1 I = FUN(1, 0) PRINT *, I, J END

9.19.2 Pointers received in the C languageIf a pointer address is returned as the function result or COMMON when the pointer is received in the C program, the result may beincorrect because such an operation is unsupported and the optimization is performed without considering it. (Refer to example 4)

In this case, using the intrinsic function LOC to store the address in the pointer will produce a correct result. (Refer to example 5)

Example4: Incorrect return of pointer

int *my_addr_(int *p){

- 279 -

return p;}

INTEGER(KIND=4) I,IPPOINTER(PTR,IP)INTEGER(KIND=8) MY_ADDRI = 2PTR = MY_ADDR(I)IP = 1PRINT *, IEND

Example5: Correct pointer usage

INTEGER(KIND=4) I,IPPOINTER(PTR,IP)I = 2PTR = LOC(I)IP = 1PRINT *, IEND

- 280 -

Chapter 10 Fortran ModulesA module reference is a USE statement specifying the module name.

This chapter describes items related to compiling input files that have modules, modules references and intrinsic modules references.

10.1 Compilation of Modules and Module ReferencesNote the following:

The object file of a module is created in the current directory. Its filename is the source filename with the suffix replaced by the suffix .o.

When a module is compiled, an information file is created by the compiler. The name of the file is the module name with the suffix .modappended. This information file is needed when compiling a program that contains a USE statement for the module. The .mod informationfile is created according to the following rules:

- If the compiler option -M is specified, it is created in the directory specified in the -M option.

- If the compiler option -M is not specified, it is created in the directory from which frtpx is being executed.

When a program that contains a USE statement is compiled, the corresponding information file is searched for in the following order:

- A directory named in the -M option (if specified).

- The directory from which frtpx is being executed.

- A directory named in the -I option (if specified).

When linking a program containing a module reference, you must specify the object file of the corresponding module.

Example1:

The module and the module reference:

! file name : a.f90

MODULE MOD INTEGER::VALUE=1END MODULE

PROGRAM MAIN USE MOD PRINT *,VALUEEND PROGRAM

Compile command :

$frtpx a.f90

Example2:

The module and the module reference appear in different files:

! file name : a.f90


! file name : b.f90

PROGRAM MAIN USE MOD PRINT *,VALUEEND PROGRAM

Compile command :

- 281 -

$ frtpx a.f90 -c$ frtpx b.f90 a.o

or

$ frtpx a.f90 -c$ frtpx b.f90 -c$ frtpx a.o b.o

Example3:

The module and two references to the module reference appear in three different files:

! file name : a.f90


! file name : b.f90

PROGRAM MAIN USE MOD VALUE=VALUE+1 CALL EXTERNAL_SUBEND PROGRAM

! file name : c.f90

SUBROUTINE EXTERNAL_SUB USE MOD PRINT *,VALUEEND SUBROUTINE

Compile command:

$ frtpx a.f90 -c$ frtpx b.f90 -c$ frtpx c.f90 a.o b.o

or

$ frtpx a.f90 -c$ frtpx b.f90 -c$ frtpx c.f90 -c$ frtpx a.o b.o c.o

Example4:

The module and the module reference appear in different files using the compiler options -I, and -M:

! file name : a.f90

MODULE MOD1 INTEGER::VALUE=1END MODULE

! file name : b.f90

MODULE MOD2 USE MOD1 INTEGER::VALUE2=2CONTAINS

SUBROUTINE MOD2_SUB VALUE=VALUE2+1

- 282 -

END SUBROUTINEEND MODULE

Compile command:

$ frtpx a.f90 -c

Note:

The information file mod1.mod for module mod1 is generated in the current directory.

$ frtpx a.f90 -c -Mdirectory

Note:

The information file mod1.mod for module mod1 is generated in the directory specified by the -M option argument.

$ frtpx b.f90 -c -Idirectory

Note:

1. The information file mod1.mod for module mod1 is searched for in the directory specified by the -I option argument or in thecurrent directory.

2. The information file mod2.mod for module mod2 is generated in the current directory.

$ frtpx b.f90 -c -Iinput_directory -Moutput_directory

Note:

1. The information file mod1.mod for module mod1 is searched for in the directory specified in the -I option argument, the directoryspecified by the -M option argument, or in the current directory.

2. The information file mod2.mod for module mod2 is generated in the directory specified by the -M option argument.

10.2 Recompiling a Program that Contains a Reference to a Modulethat Has Changed

In either of the following cases, a program unit that contains a module reference must be recompiled.

- An entity with the PUBLIC attribute in the module specification part is changed

- The characteristics of a module procedure with the PUBLIC attribute are changed

10.3 Restrictions on ModulesWhen the following compiler option is specified for compile a module, you must specify the same option for compile the other programthat including the module reference.

- -AA option

- -AU option

- -Kcommonpad[=N] option

- -Karray_subscript option

When a module is compiled with compiler option -X9 or -X03, the other program including the module reference also has to be compiledwith compiler option -X9 or -X03.

When a module is compiled with compiler option -X6, the other program including the module reference also has to be compiled withcompiler option -X6.

When a module is compiled with compiler option -X7, the other program including the module reference also has to be compiled withcompiler option -X7.

- 283 -

10.4 Intrinsic ModulesThe intrinsic modules are provided in the processor. There are standard intrinsic module and non-standard intrinsic modules in the intrinsicmodules. The Fortran environment module (ISO_FORTRAN_ENV), IEEE module (IEEE_EXCEPTIONS, IEEE_ARITHMETIC andIEEE_FEATURES), module for interoperable with C program (ISO_C_BINDING) are standard intrinsic modules. TheSERVICE_ROUTINES and OMP_LIB modules are the non-standard intrinsic module. See 6.6.6 Service Routines for information aboutSERVICE_ROUTINES modules.

Be careful to the following when the intrinsic module is referenced.

- When the compiler option -AU is specified, the language entities of intrinsic modules must be lowercase letter.

10.4.1 COMPILER_OPTIONS intrinsic module functionCompiler options information on the function result is the same as compiler options information with compiler option -Q or -Nlst. Consider1200 characters or more for character strings of the function result on programming.

10.4.2 COMPILER_VERSION intrinsic module functionVersion information on the function result is the same as information on the Fortran compiler when compiler option -V is specified.Consider 90 characters or more for character strings of the function result on programming.

- 284 -

Chapter 11 Mixed Language ProgrammingThis chapter describes the specification and notes when linking a Fortran program and a C/C++ program.

The description of C program on this chapter is applied C++ program. In linking with the C++ program, see Section "Notes on InterlanguageLinkage" of the "C++ User's Guide" in addition to this chapter.

11.1 Summary of Mixed Language ProgrammingA Fortran program and a C program can be linked statically and executed.

For the C and C++ languages, there are two types of compile commands, one is the cross compile command, and the other is the owncompile command.

Programming Language Cross Compiler Own Compiler

C fccpx fcc

C++ FCCpx FCC

Summary of compiling and linking a Fortran program and a C program are shown in the following figure.

11.2 Features for Mixed Language ProgrammingThis system has Fortran standard and FUJITSU Fortran extension specifications for mixed language programming.

- 285 -

The following Fortran standard specifications for mixed language programming are offered.

- Intrinsic module ISO_C_BINDING

- BIND statement, BIND attribute specifier, Language binding specifier, Procedure language binding specifier

- VALUE statement, VALUE attribute specifier

The following FUJITSU Fortran extension specifications for mixed language programming are offered.

- VAL intrinsic function, %val(a) specifies that an actual argument

- CHANGEENTRY statement

- $pragma specifier

The following features are supported by compiler option.

- The compiler options -ml and -mldefault to change how to make external procedure. See Section "11.3 Rules for Processing FortranProcedure Names".

- The compiler option -Az to change how to call the value of character type for the argument. See Section "11.13 Notes of the CompilerOption -Az".

11.2.1 Passing Arguments by ValueBy default, actual arguments of external procedure reference are called by address. See Section "11.7 Passing and Getting Arguments".In the following cases, the actual argument is passed by value.

- The intrinsic function VAL(a) or %val(a) for actual argument is specified.

- The VALUE attribute is specified in a procedure interface.

11.2.1.1 Passing Arguments by Value in Procedure InterfaceWhen the VALUE attribute is specified in procedure interface, the argument is passed by value.

See Section "11.2.2 Getting Arguments by Value" for information about the actual argument that can be associated.

Example: the VALUE attribute in procedure interface

INTERFACE SUBROUTINE CSUB(ARG1,ARG2) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT),VALUE::ARG1 INTEGER(C_INT) ::ARG2 END SUBROUTINE CSUBEND INTERFACECALL CSUB(2,3) ! The first argument is passed by value, !the second argument is called by address.END

11.2.1.2 Passing Arguments by Value in Intrinsic FunctionWhen the compiler option -Nobsfun is specified, the function VAL(a) and %val(a) are intrinsic functions. The intrinsic function VAL(a)and %val(a) specifier can specify only for actual argument of external procedure reference. The actual argument passed by value must bescalar expression of type one-byte integer, two-byte integer, four-byte integer, eight-byte integer, one-byte logical, two-byte logical, four-byte logical, eight-byte logical, single-precision real or double-precision real.

Example: intrinsic function VAL()

CALL CSUB(VAL(2),3) ! first argument passed by value, ! second argument passed by addressEND

- 286 -

11.2.2 Getting Arguments by ValueBy default, dummy arguments of external procedure are got by address. See Section "11.7 Passing and Getting Arguments". The VALUEattribute is specified for a dummy argument of an external procedure, the Fortran program gets the dummy arguments by value. TheVALUE attribute can specify by the VALUE statement or attribute specifier in a type declaration statement.

When the dummy argument has the Quadruple-precision complex, the character type of procedure without procedure language bindingspecifier, the derived type whose size is 16 bytes or less, the derived type who has POINTER attribute in component, the derived typewho has ALLOCATABLE attribute in component or OPTIONAL attribute, the address of the variable is always received.

Example: VALUE statement

SUBROUTINE CSUB(I,K)VALUE :: I ! first argument passed by value, ! second argument passed by addressRRINT *, I,K END

11.2.3 CHANGEENTRY StatementThe CHANGEENTRY statement changes rules for processing Fortran procedure names combine with -ml compiler option. See Section"11.3 Rules for Processing Fortran Procedure Names" for details.

The external procedure name must not have the procedure language binding specifier.

Example: CHANGEENTRY statement

CHANGEENTRY :: CSUBCALL CSUB()END

11.2.4 $pragma Specifier$pragma specifier has the following form and can be specified like a comment.

!$pragma c (prc)

prc must be an external procedure name. The $pragma specifier specifies that the procedure prc is coded by C program and changes rulesfor processing procedure names. See Section "11.3 Rules for Processing Fortran Procedure Names".

The external procedure name of $pragma specifier must not have the procedure language binding specifier.

Example: $pragma specifier

EXTERNAL CSUB !$pragma C(CSUB) :CALL CSUB ! CSUB is treated as a calling procedure in C program.END

11.2.5 Intrinsic Module ISO_C_BINDINGThe following specifications are offered by intrinsic module ISO_C_BINDING.

- Named constants for interoperable with C program

- Types for interoperable with C program

- Intrinsic procedures for interoperable with C program

11.2.5.1 Named Constants by Intrinsic Module ISO_C_BINDINGThe following named constants are offered by intrinsic module ISO_C_BINDING.

- Named constants for kind type parameter

- 287 -

- Named constants for names of C program character

- Named constants for null pointer

1. The value of named constants by intrinsic module ISO_C_BINDING and C program types correspond in "Table 11.1 Types of Cprogram and, types and kind type parameters of Fortran correspond". If data is passed between Fortran and C programs, type of Cprogram and, type and kind type parameter of Fortran must correspond. If data type is character, the value of type length parametermust be 1.

Table 11.1 Types of C program and, types and kind type parameters of Fortran correspond

Named constants Value C program types Type specifiers of Fortran

C_SHORT 2 short int INTEGER

C_INT 4 int INTEGER

C_LONG 8 long int INTEGER

C_LONG_LONG 8 long long int INTEGER

C_SIGNED_CHAR 1 signed char / unsigned char INTEGER

C_SIZE_T 8 size_t INTEGER

C_INT8_T 1 int8_t INTEGER




C_INT_LEAST8_T 1 int_least8_t INTEGER




C_INT_FAST8_T 1 int_fast8_t INTEGER




C_INTMAX_T 8 intmax_t INTEGER

C_INTPTR_T 8 intptr_t INTEGER

C_FLOAT 4 float REAL

C_DOUBLE 8 double REAL

C_LONG_DOUBLE 16 long double REAL

C_FLOAT_COMPLEX 4 float _Complex COMPLEX

C_DOUBLE_COMPLEX 8 double _Complex COMPLEX

C_LONG_DOUBLE_COMPLEX

16long double _Complex COMPLEX

C_BOOL 1 _Bool LOGICAL

C_CHAR 1 char CHARACTER

Example: Using of named constant of kind type parameter by intrinsic module ISO_C_BINDINGUSE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT

INTEGER(C_INT)::J,K,LINTERFACE SUBROUTINE SUB(J,K,L) BIND(C)

- 288 -

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT)::J,K,L END SUBROUTINEEND INTERFACEJ = 10K = 20L = 30CALL SUB(J,K,L)END

C program:

#include <stdio.h>void sub(j,k,l)int *j,*k,*l;{printf ("J:%d K:%d L:%d \n",*j,*k,*l) ;}

2. Named constants of C program character by intrinsic module ISO_C_BINDING, and the meaning are follows.

Named constants C program characters

C_NULL_CHAR null character

C_ALERT alert

C_BACKSPACE backspace

C_FORM_FEED form feeds

C_NEW_LINE new line

C_CARRIAGE_RETURN carriage return

C_HORIZONTAL_TAB horizontal tab

C_VERTICAL_TAB horizontal tab

The type of named constants is character type and length type parameter is 1.

Example: Using of named constant of C program character by intrinsic module ISO_C_BINDING

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_NULL_CHARCHARACTER(LEN=5)::STST='ABCD'//C_NULL_CHAR ! Append null character:END

3. Named constants of null pointer by intrinsic module ISO_C_BINDING, the meaning and type are follows.

Named constants Meaning Type

C_NULL_PTR NULL C_PTR

C_NULL_FUNPTR null pointer to function C_FUNPTR

Example: Using of named constant of null pointer by intrinsic module ISO_C_BINDING

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_PTR,C_NULL_PTRINTERFACE SUBROUTINE SUB(PTR) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_PTR TYPE(C_PTR),VALUE::PTR END SUBROUTINEEND INTERFACETYPE(C_PTR)::PTRPTR=C_NULL_PTR ! Set NULL

- 289 -

CALL SUB(PTR)END

11.2.5.2 Types by Intrinsic Module ISO_C_BINDINGC_PTR type and C_FUNPTR type are offered by intrinsic module ISO_C_BINDING.

C_PTR type is a pointer type of C program. C_FUNPTR type is a function pointer type of C program. C_PTR and C_FUNPTR are derivedtype with private component.

11.2.5.3 Intrinsic Procedures by Intrinsic Module ISO_C_BINDINGIntrinsic functions C_LOC, C_FUNLOC, C_ASSOCIATED, C_SIZEOF and, Intrinsic subroutines C_F_POINTER,C_F_PROCPOINTER are offered by intrinsic module ISO_C_BINDING.

C_LOC gets pointer (C_PTR type) of C program corresponding to the argument.

C_FUNLOC gets function pointer (C_FUNPTR type) of C program corresponding to the argument.

C_ASSOCIATED compares pointer or function pointer of C program corresponding to the argument.

C_F_POINTER assigns data pointer of C program to pointer of Fortran.

C_F_PROCPOINTER assigns function pointer of C program to procedure pointer of Fortran.

C_SIZEOF gets size of the argument.

See online manual intrinsic(3) for details of these intrinsic procedures.

Example: Using intrinsic procedure by intrinsic module ISO_C_BINDING

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_PTR,C_INT,C_LOCINTERFACE SUBROUTINE SUB(PTR) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_PTR TYPE(C_PTR),VALUE::PTR END SUBROUTINEEND INTERFACETYPE(C_PTR)::PTRINTEGER(C_INT),TARGET::TARGETTARGET=2PTR=C_LOC(TARGET) ! Set C program pointer of variable TARGETCALL SUB(PTR)END

C program:

#include <stdio.h>void sub(k)int *k;{printf ("K=%d \n",*k) ;}

11.2.6 BIND Statement, BIND Attribute Specifier, Language BindingSpecifier and Procedure Language Binding Specifier

When a language binding specifier is specified, data can interoperate between Fortran and C program.

When a procedure language binding is specified, procedure can interoperate between Fortran and C program.

The language binding specifier and procedure language binding specifier are following syntax.

BIND ( C [, NAME = scalar character constant expression] )

- 290 -

External name of language entity is decided by language binding specifier or procedure language binding specifier. The external name bythe specifier is called a binding label.

If the language binding specifier or procedure language binding specifier is specified, the external name is the follows. However, valueof NAME specifier is discarding leading and trailing blanks.

- If NAME specifier is not specified or length of NAME specifier value is zero, the binding label is the same as the name of the entityusing lower case letters.

- If NAME specifier value has nonzero length, the binding label has the value. The case of letters in the binding label is significant.

If the language binding specifier of a BIND statement or attribute specifier in type declaration statement is specified, a variable or commonblock linkage interoperates with an external variable of C program.

If the language binding specifier is specified, the data entity has BIND attribute.

A procedure language binding specifier can specify in SUBROUTINE statement, FUNCTION statement or ENTRY statement, and theprocedure interoperates with C program.

Example: Specify language binding specifier and procedure language binding specifier

MODULE MOD INTEGER,BIND(C,NAME='gvar01')::VAR INTEGER,BIND(C )::GVAR02END MODULE

USE MODUSE,INTRINSIC::ISO_C_BINDING,ONLY:C_INTINTERFACE SUBROUTINE SUB() BIND(C,NAME='Sub_proc') END SUBROUTINEEND INTERFACEINTEGER(C_INT)::K1,K2COMMON /COM/ K1,K2BIND(C,NAME='gvar03'):: /COM/VAR=1GVAR02=2K1=3K2=4CALL SUBEND

C program

#include <stdio.h>struct {int k1; int k2;} gvar03;int gvar01;int gvar02;

void Sub_proc(void){printf ("gvar01=%d \n",gvar01) ;printf ("gvar02=%d \n",gvar02) ;printf ("gvar03.k1=%d \n",gvar03.k1) ;printf ("gvar03.k2=%d \n",gvar03.k2) ;}

11.2.6.1 BIND StatementA BIND statement specify BIND attribute, and the statement has the following form.

Language binding specifier[ :: ] binding-entity-list

The binding-entity is a variable name or /common-block-name/. If a variable name is specified, the BIND statement must appear indeclaration part of module.

- 291 -

If a NAME specifier specify in language binding specifier, then binding-entity-list must contain exactly one binding-entity.

Example: Interoperates with C program by BIND statement

MODULE MOD USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT)::N INTEGER(C_INT)::K1,K2 COMMON /COM/ K1,K2 BIND(C):: N,/COM/END MODULE

USE MODINTERFACE SUBROUTINE SUB() BIND(C) END SUBROUTINEEND INTERFACEN=1K1=2K2=3CALL SUBEND

C program

#include <stdio.h>struct {int k1; int k2;} com;int n;

void sub(void){printf ("n=%d \n",n) ;printf ("com.k1=%d \n",com.k1) ;printf ("com.k2=%d \n",com.k2) ;}

11.2.6.2 BIND Attribute by Attribute Specifier of Type Declaration StatementA language binding specifier is specified in attribute specifier of type declaration statement, BIND attribute is specified.

A language binding specifier of type declaration statement can specify in declaration part of module. If a NAME specifier specify inlanguage binding specifier, then variable list must contain exactly one variable.

Example: Specify language binding specifier of type declaration statement.

MODULE MOD INTEGER,BIND(C,NAME='global01')::VAR INTEGER,BIND(C ):: GLOBAL02,GLOBAL03END MODULE

USE MODINTERFACE SUBROUTINE SUB() BIND(C) END SUBROUTINEEND INTERFACEVAR=1GLOBAL02=2GLOBAL03=3CALL SUBEND

C program

#include <stdio.h>int global01;

- 292 -

int global02;int global03;

void sub(void){printf ("global01=%d \n",global01) ;printf ("global02=%d \n",global02) ;printf ("global03=%d \n",global03) ;}

11.2.6.3 Procedure Language Binding SpecifierA procedure language binding specifier specify following dummy argument list in SUBROUTINE statement, FUNCTION statement orENTRY statement. When the procedure language binding is specified and the procedure doesn't have dummy argument list, parenthesesenclose the dummy argument are not omissible.

Example: Specify procedure language binding specifier

SUBROUTINE SUB(K) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT)::K PRINT *,'K=',K END SUBROUTINE FUNCTION IFUN(K) BIND(C,NAME='fun') RESULT(R) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT)::R,K R=K END FUNCTION

C program

#include <stdio.h>void sub(int *);int fun(int *) ;int foo(void){int n,k;k=2 ;sub(&k) ;n=fun(&k) ;printf (" n= %d \n",n) ;return 0 ;}

11.2.7 VALUE Statement, VALUE Attribute SpecifierA VALUE statement or VALUE attribute specifier of type declaration statement is specified, the dummy argument has VALUE attribute.If program calls procedure of VALUE attribute dummy argument, the procedure must have an explicit interface. When a dummy argumenthas VALUE attribute, the dummy argument is passed by value without follows:

- A character type of procedure without procedure language binding

- A component has POINTER attribute in derived type

- A component has ALLOCATABLE attribute in derived type

- OPTIONAL attribute

Example: Specify a VALUE attribute and VALUE statement

SUBROUTINE SUB(K) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT),VALUE::K PRINT *,'K=',K

- 293 -

END SUBROUTINE! USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTERFACE SUBROUTINE FOO(K1,K2) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT INTEGER(C_INT),VALUE::K1 INTEGER(C_INT) ::K2 VALUE::K2 END SUBROUTINE END INTERFACE INTEGER(C_INT)::N1,N2 N1=1 N2=2 CALL FOO(N1,N2)END

C program

void sub(int);int foo(int k1,int k2){sub(k1) ;sub(k2) ;return 0 ;}

11.2.8 Interoperability of Derive type and C language structure typeIn case Fortran derived type is interoperable with structure type of C language, the following conditions should be met:

- Fortran derived type must have BIND Attribute.

- Fortran derived type must not be sequence type.

- Fortran derived type must not have Type parameter.

- Fortran derived type must not have EXTENDS attribute.

- Fortran derived type must not have type bound procedure part.

- Each component of Fortran derived type must be interoperable type as well as type parameter.

- Each component of Fortran derived type must not be pointer or allocatable.

- Fortran derived type must have same number of components as that of C language structure type.

- Each component of Fortran derived type must have component type which supports structure type of C language, interoperable typeand type parameter.

11.3 Rules for Processing Fortran Procedure NamesIn the object, the Fortran procedure name processed from source name. If a Fortran procedure name is called from C program, the namemust conform to the Fortran name rules.

The following table lists the rules for processing Fortran procedure names.

Table 11.2 Rules for processing Fortran procedure names

Procedure name Processing name

Main program name MAIN__

External procedure name external-procedure-name_

Block data program unit The BLOCK DATA program with a name block-data-program-unit-name_

- 294 -

Procedure name Processing name

The BLOCK DATA program without a name _BLKDT__

Startup routine main

Notes:

1. All procedure names are uniformly lowercase if compiler option -AU is not specified.

2. The startup routine is the routine to activate the Fortran system.

If compiler option -AU is specified, the processing names are same spell in source program.

The processing name of the procedure with procedure language binding specifier is the binding label.

Rules for Processing Fortran Procedure Names can be changed by compiler option -mldefault.

When the compiler option -mldefault is specified, the procedures with procedure language binding specifier, the procedures specified inCHANGEENTRY statement and the external procedures name specified in $pragma specifier are not effective.

The following table lists the rules for processing Fortran external procedure names when the -mldefault option is specified.

Table 11.3 Rules for processing Fortran external procedure names when the -mldefault option is specified

Compiler option -mldefault Processing name

-mldefault=cdecl external-procedure-name

-mldefault=frtexternal-procedure-name_

Not specified

Compiler option -ml changes rules for processing Fortran external procedure names of external procedure name statement. The followingtable lists rules for processing Fortran external procedure names.

Table 11.4 Rules for processing Fortran external procedure names specified in CHANGEENTRY statement

Compiler option -ml Processing name

-ml=cdecl external-procedure-name

-ml=frt external-procedure-name_

Not specified in accordance with specifying -mldefault option

The following table lists rules for processing Fortran external procedure names specified in $pragma specifier.

Table 11.5 Rules for processing Fortran external procedure names in $pragma specifier

$pragma specifier Processing name

c (prc) external-procedure-name

11.4 Correspondence of type with CIf data is passed between Fortran and other programs, data attributes must correspond. See Section "5.1 Internal Data Representation" forFortran data types.

When data is passed between Fortran and C programs, it is necessary to make the type correspond.

See Section "Table 11.1 Types of C program and, types and kind type parameters of Fortran correspond" for the type that can be sharedwhen the BIND attribute is specified.

The following table lists data type that Fortran and C when BIND attribute is not specified correspond.

Table 11.6 Data type that A and B correspond

Data Type Fortran C

One-byte logical LOGICAL(1) L1 unsigned char L1 ;

Two-byte logical LOGICAL(2) L2 short int L2 ;

- 295 -

Data Type Fortran C

Four-byte logical LOGICAL(4) L4 int L4 ;

Eight-byte logical LOGICAL(8) L8 long long int L8 ;

One-byte integer INTEGER(1) I1 signed char I1 ;

Two-byte integer INTEGER(2) I2 short int I2 ;

Four-byte integer INTEGER(4) I4 int I4 ;

Eight-byte integer INTEGER(8) I8 long long int I8 ;

Single -precision real REAL(4) R4 float R4 ;

Double-precision real REAL(8) R8 double R8 ;

Quadruple-precision real REAL(16) R16 long double R16 ;

Single -precision complex COMPLEX(4) C8struct{ float r,i;} C8 ;

Double-precision complex COMPLEX(8) C16struct{ double r,i;} C16 ;

Quadruple-precision complex COMPLEX(16) C32struct{ long double r,i; } C32 ;

Character CHARACTER(10) S char S [10] ;

Derived type

TYPE TAGINTEGER I4REAL(8) R8END TYPE TAGTYPE(TAG)::D

struct tag{ int i4;double r8; } d;

11.5 Calling ProceduresIf a Fortran procedure name is called from a C program or C function name is called from a Fortran program, the processing name mustconfirm. See Section "11.3 Rules for Processing Fortran Procedure Names" for processing Fortran procedure names.

The example of the calling procedure is shown as follows.

Example 1: Calling C function from Fortran program with the procedure language binding specifier.

Fortran program:

INTERFACE SUBROUTINE SUB() BIND(C) END SUBROUTINEEND INTERFACECALL SUB( )END

C program:

#include <stdio.h>void sub( ){printf ("%s\n","** sub **");}

The external name of Fortran calling procedure SUB is the binding label sub. Therefore, function name of C is sub.

Example 2: Calling Fortran procedure with the procedure language binding specifier from C program

C program:

- 296 -

void sub( );int c_( ){sub( );return 0;}

Fortran program:

SUBROUTINE SUB() BIND(C)PRINT *,'** SUB **'END

The external name of Fortran subroutine SUB is the binding label sub. Therefore, subroutine SUB is able to interoperable.

Example 3: Calling C function from Fortran program without the procedure language binding specifier

Fortran program:

EXTERNAL SUBCALL SUB( )END

C program:

#include <stdio.h>void sub_( ){printf ("%s\n","** sub **");}

In Fortran program, calling SUB are processed "sub_", so C function name must be "sub_".

Example 4: Calling Fortran procedure without the procedure language binding specifier from C program

C program:

void sub( );int c_( ){sub( );return 0;}

Fortran program:

SUBROUTINE SUB( )CHANGEENTRY :: SUBPRINT *,'** SUB **'END

In Fortran program, specify procedure name in CHANGEENTRY statement and specify compiler option -ml=cdecl to confirmprocessing procedure names.

11.5.1 Passing Control First to a C ProgramIf the program to which control is first passed is in C, the function name must be MAIN__. Do not use main. However, if the compileroption -mlcmain=main is specified, main can be used for the function name.

Example 1: Passing control first to a C program:

C program: a.c

void sub_( );int MAIN__( ){

- 297 -

sub_( );return 0;}

Fortran program: b.f95

SUBROUTINE SUB( )PRINT *,'SUB'END

Compile, link and execute:

$ fccpx -c a.c$ frtpx a.o b.f95$ ./a.outSUB$

Example 2: Passing control first to a C program (function name is main)

C program: a.c

void sub_( );int main( ){sub_( );return 0;}

Fortran program: b.f95

SUBROUTINE SUB( )PRINT *,'SUB'END

Compile, link and execute:

$ fccpx -c a.c$ frtpx -mlcmain=main a.o b.f95$ ./a.outSUB$

11.5.2 Calling Standard C LibrariesWhen a standard C library is called from the Fortran program, there are two methods.

- The method of the specification of the procedure language binding specifier

- The method of no specification of the procedure language binding specifier

11.5.2.1 Method of the Specification of the Procedure Language Binding SpecifierA standard library of C can be called from the Fortran program.

Example: Calling a standard C library with the procedure language binding specifier.

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_NULL_CHARINTERFACE FUNCTION STRCMP(STR1,STR2) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_INT,C_CHAR INTEGER(C_INT)::STRCMP CHARACTER(KIND=C_CHAR),DIMENSION(*)::STR1,STR2 END FUNCTIONEND INTERFACE

- 298 -

CHARACTER(LEN=5)::ST1,ST2ST1='ABCD'//C_NULL_CHARST2='ABCD'//C_NULL_CHARIF (STRCMP(ST1,ST2)==0)PRINT *,"same string"END

In this example, the actual argument of the character type scalar corresponds to the dummy argument of the character type array.

Compile and execute:

$ frtpx a.f95 $ ./a.outsame string$

11.5.2.2 Method of no Specification of the Procedure Language Binding SpecifierTo calling standard C libraries from Fortran program directory, you must do following:

- Specify the standard C library name in CHANGEENTRY statement

- Specify compiler option -ml=cdecl

See Section "11.3 Rules for Processing Fortran Procedure Names", and be careful to passing argument. See Section "11.7 Passing andGetting Arguments" for details.

Example: Calling a standard C library without the procedure language binding specifier.

Fortran program:a.f95

INTEGER :: STRCMPCHANGEENTRY :: STRCMPCHARACTER(LEN=5) :: ST1,ST2ST1 = 'ABCD\0' ; ST2 = 'ABCD\0' ! Compiler option -Ae need to be specified.IF (STRCMP(ST1,ST2)==0) PRINT *," same string "END

Compile and execute:

$ frtpx a.f95 -ml=cdecl -Ae$ ./a.out same string$

11.6 Passing and Receiving Function ValuesThis section describes passing and receiving function values between Fortran and C program without the procedure language bindingspecifier. See "Table 11.1 Types of C program and, types and kind type parameters of Fortran correspond" for the type that can be sharedwhen the procedure language binding is specified.

11.6.1 Passing Fortran Function Values to C without the ProcedureLanguage Binding Specifier

The following table lists the methods for calling Fortran procedures from C programs for each type of function result that a Fortran programreturns.

Fortran function value How C accepts values Example of values returned by C

INTEGER(1) signed char

res = sub_();INTEGER(2) short int

INTEGER(4) int

- 299 -

Fortran function value How C accepts values Example of values returned by C

INTEGER(8) long long int

REAL(4) float

REAL(8) double

REAL(16) long double

LOGICAL(1) unsigned char

LOGICAL(2) short int

LOGICAL(4) long int

LOGICAL(8) long long int

COMPLEX(4)

structureCOMPLEX(8)

COMPLEX(16)

CHARACTER void sub_(&res, len); (*1)

Derived type structure res = sub_();

sub_: Called Fortran procedure name

res: Area for the function value returned. See Section "11.4 Correspondence of type with C" about the corresponding type

len: Character length of the function type. Then, Fortran receives character length as a value of eight byte integer type.

*1) The area of the return value of the procedure without the procedure language binding specifier is used the area of the actualargument. In this case, the address of the return value is put in the first argument and the length of the return value is put in the secondargument. And the actual argument is put after the second arguments.

Example 1: Acceptance of an integer function value:

C program:

#include <stdio.h>int ifun_();int main( ){ int i; i = ifun_(); printf("%d\n",i); return 0;}

Fortran program:

FUNCTION IFUN()INTEGER IFUNIFUN = 100END

Example 2: Acceptance of a double-precision real function value:

C program:

#include <stdio.h>double r8fun_() ;int main( ){ double r8; r8 = r8fun_() ; printf("%f\n",r8);

- 300 -

return 0;}

Fortran program:

FUNCTION R8FUN()REAL(8) R8FUNR8FUN= 12.2D+0END

Example 3: Acceptance of a character function value:

C program:

#include <stdio.h>void cfun_(char *, long long int);int main( ){ char cha [11] ; cfun_(cha,10); printf("%s\n",cha); return 0;}

Fortran program:

FUNCTION CFUN()CHARACTER(LEN=*) CFUNCFUN = '1234567890'END

To accept the value of character function CFUN, set the address of the return area to the first argument, and set the length of functionto the second argument. Then, Fortran receives length of character function CFUN as a value of eight byte integer type.

Example 4: Acceptance of a derived type function value:

C program:

#include <stdio.h>struct tag{ int i; double fr8; };struct tag dfun_( );int main( ){ struct tag d; d = dfun_( ); printf("%d %f\n",d.i, d.fr8); return 0;}

Fortran program:

FUNCTION DFUN()TYPE TAG SEQUENCE INTEGER(4) I REAL(8) FR8END TYPETYPE (TAG) ::DFUNDFUN%I = 10DFUN%FR8 = 1.2D+0END

- 301 -

11.6.2 Receiving Function Values from C without the Procedure LanguageBinding Specifier

The following table lists the methods for calling C functions from Fortran programs without the procedure language binding Specifier foreach type of function result that a C program returns.

Specifier for each type of function result that a C program returns.

C function value How Fortran accepts value Example of values returned by Fortran

void Not accepted CALL SUB()

signed char INTEGER(1)

RES = SUB()

short int INTEGER(2)

int INTEGER(4) or LOGICAL(4)

long long int INTEGER(8)

float REAL(4)

double REAL(8)

long double REAL(16)

structure Derived type

SUB: Calling C function processed in sub_. See Section "11.3 Rules for Processing Fortran Procedure Names".

RES: Area for the function value returned.

Example 1: Acceptance of an integer function value:

Fortran program:

INTEGER(4) I,IFUNI = IFUN()PRINT *,IEND

C program:

int ifun_(){ return 100;}

Example 2: Acceptance of a derived type function value:

Fortran program:

TYPE TAG INTEGER(4) REAL(8) FR8END TYPETYPE (TAG) ::DFUN,DD = DFUN()PRINT *,D%I,D%FR8END

C program:

struct tag{int i; double fr8; } ;struct tag dfun_(){ struct tag d ;

- 302 -

d.i=100 ; d.fr8=1.2 ; return(d) ;}

11.7 Passing and Getting ArgumentsThe passing and receiving arguments between procedures are the address in the default. If arguments are passed and received from Fortranprogram, see Section "11.2.1 Passing Arguments by Value".

If a character type argument is passed, the argument length as a value of eight byte integer type is also passed. The actual argument thatshows the argument length exists only by the number of arguments of the character type be after all the specified arguments.

However, the argument length is not passed when the argument is the procedure with the procedure language binding specifier or VALUEattribute. The following sections explain how arguments pass data without the procedure language binding specifier.

11.7.1 Passing Arguments from Fortran with the Procedure LanguageBinding Specifier to C

Because Fortran passes actual arguments by address, the address is accepted if dummy arguments are declared in C. Specify VAL( )intrinsic function or %val( ) specifier to pass the actual argument by value.

See Section "11.2.1 Passing Arguments by Value" for details.

Example: Passing integer and real type arguments:

Fortran main program:

PROGRAM FORTRAN_MAINWRITE(*,*) '*** START ***'J=10K=20L=30CALL FUN1(J,K,L)F=10.0CALL FUN2(F)WRITE(*,*) '*** FMAIN END ***'END

C:

#include <stdio.h>void fun1_(j,k,l)int *j,*k,*l;{ printf("J:%d K:%d L:%d\n",*j,*k,*l);}

void fun2_(f)float *f;{ printf("F:%f\n",*f);}

11.7.2 Passing Arguments from C to Fortran without the ProcedureLanguage Binding Specifier

Because Fortran accepts dummy arguments by address in the default, the address is passed if actual arguments appear in C. Specify VALUEattribute for dummy argument to get the actual argument by value. See Section "11.2.2 Getting Arguments by Value" for details.

- 303 -

Example 1: Passing integer type arguments:

C program (MAIN__):

#include <stdio.h>int MAIN__(){ int fun_(); int i,j,k; i=10; j=20; k=fun_(&i,&j); printf(" k:%d\n",k); return 0;}

Fortran program:

INTEGER FUNCTION FUN(X,Y)INTEGER(4) X,Y

WRITE(*,*) '**** FUNC1 **** 'FUN =X+YEND

Example 2: Passing character type arguments:

C program (MAIN__):

#include <stdio.h>int MAIN__( ){ void fun_(char *, long long int, char *, long long int); char res[3],ch; ch = 'o'; res[2] = 0; fun_(res,2,&ch,1); printf(" res:%s\n",res); return 0;}

Fortran:

CHARACTER(LEN=*) FUNCTION FUN(CH)CHARACTER(LEN=*) CHFUN = CH//'k'END

Note:

Character type function FUN has one character type argument. If this function is called from a C program, four arguments are passed.The first argument is the address of the area for the function value. The second argument is the length of the function value. The thirdargument is the address of the argument ch. The fourth argument is the length of ch. Then, Fortran receives result length and argumentlength of character type function FUN as a value of eight byte integer type.

11.8 Passing Using External Variables- The method of the specification of the BIND attribute.

- The method of no specification of the BIND attribute.

- 304 -

11.8.1 Passing External Variables with BIND AttributeWhen COMMON block or module variable with BIND attribute, this COMMON block or variable can associate to the external variable.The external name is the binding label of the variable or COMMON block.

Example 1: Associating the external variable with BIND attribute

MODULE MOD INTEGER,BIND(C)::INT_VARIABLE ! Associating the external variable int_variable in C program END

Example 2: Passing the named COMMON block with BIND attribute

Fortran program:

INTERFACE SUBROUTINE SUB() BIND(C) END SUBROUTINE SUBEND INTERFACEINTEGER I,JCOMMON /EXT/I,JBIND(C)::/EXT/I = 1J = 1CALL SUB( )END

C program:

#include <stdio.h>struct tag { int i,j; } ext;void sub( ){printf("i=%d j=%d\n",ext.i,ext.j);}

Passed COMMON block EXT with BIND attribute.

11.8.2 Passing External Variables without BIND attributeWithout BIND attribute, the data in Fortran COMMON block can be passed as the external structure type variable in C program.

A COMMON block name of Fortran and an external variable of C program should be the following correspondences.

Specification of COMMON block External variable name of structure of C

Named COMMON block COMMON block name_

Unnamed COMMON block _BLNK__

The COMMON block name is uniformly lowercase if compiler option -AU is not specified. If compiler option -AU is specified, theCOMMON block name is same spell in source program.

Example 1: Named common block without BIND attribute

Fortran program:

COMMON /EXT/I,JI=1J=2CALL SUB()END

C program:

- 305 -

#include <stdio.h>struct tag {int i,j;} ext_;void sub_(){ printf("i=%d j=%d\n",ext_.i,ext_.j);}

Note:

To reference common block EXT, change the name to lowercase and add an underscore to the name.

Example 2: Blank common block without BIND attribute

Fortran program:

COMMON //X,YX=1.0Y=2.0CALL SUB()END

C program:

#include <stdio.h>struct tab {float x,y;}_BLNK__ ;void sub_(){printf("x=%f y=%f\n",_BLNK__.x, _BLNK__.y) ;}

11.9 Passing Using FilesIf data is passed by file, the file must be closed. To close files in Fortran and C programs, use the following instructions:

Fortran C

CLOSE statement close (2)

11.10 Array Storage SequenceFortran and C differ in the order in which they store elements in arrays of two or more dimensions.

Therefore, be careful when passing arrays between Fortran and C.

The following shows the different storage sequences in Fortran and C arrays.

- 306 -

11.11 Passing a Character String to a C ProgramThe null character "\0" normally is not appended to the constants of Fortran character type. If a function such as strlen (standard C libraryfunction) is called, you must do one of the followings:

- Add the intrinsic module C_NULL_CHAR (null character) of ISO_C_BINDING to the end of character strings.

- Append a null character (CHAR(0)) to the end of constants. Then, compiler option -Ae must be specified.

- Specify the compiler option -Az when the constants of character type is specified in the argument.

Example 1: Passing and receiving the character string using BIND attribute and C_NULL_CHAR

Fortran program:

USE,INTRINSIC::ISO_C_BINDING,ONLY:C_NULL_CHARINTERFACE SUBROUTINE SUBA(STR) BIND(C) USE,INTRINSIC::ISO_C_BINDING,ONLY:C_NULL_CHAR,C_CHAR CHARACTER(KIND=C_CHAR)::STR(*) END SUBROUTINEEND INTERFACECALL SUBA('12345'//C_NULL_CHAR) ! Add null character to the end of the character stringsEND

C program:

#include <stdio.h>#include <string.h>void suba(char *str){int i,len;printf("len=%d,[%s]\n",len=strlen(str),str);for (i=0; i < len; i++){printf("0x%02x ",*str++);}putchar('\n');}

Example 2: Passing and receiving the character strings added NULL character not using BIND attribute.

Fortran program:

- 307 -

! CALL SUBA('12345') CALL SUBA('12345\0') ! Add \0 to the end of the character strings. END

C program:

#include <stdio.h>#include <string.h>void suba_(char *str){int i,len;printf("len=%d,[%s]\n",len=strlen(str),str);for (i=0; i < len; i++){printf("0x%02x ",*str++);}putchar('\n');}

Note:

Compiler option -Ae must be specified for Fortran program.

11.12 Return Value of Executable ProgramWhen the Fortran program is linked with C/C++ language program, the return value of executable program is decided as follows.

- In the following cases, the return value of Fortran is set the return value of executable program. See Section "3.6 Execution CommandReturn Values".

- The program to which control is first passed is Fortran program.

- In the following cases, the return value of C/C++ is set the return value of executable program.

- The program to which control is first passed is C/C++ program.

11.13 Notes of the Compiler Option -AzIf the compiler option -Az is to add \0(Null character) to the end of character strings. However, a part of execution is not guaranteed bythe program described by the specification before FORTRAN77.

Example: The compiler option -Az

CALL SUB('12')ENDSUBROUTINE SUB(I)INTEGER(4) IIF (I.NE.'12') PRINT *,'ok'END

Please change the CALL statement of the above-mentioned program as follows.

CALL SUB('12 ')

11.14 Restrictions on Linking with C Language ProgramsThe following restrictions and notes apply to linkage with C programs. And, see Section "9.19 Attentions for Mixed LanguageProgramming" for notes combined with the optimization functions.

- With BIND attribute, the language entities of Fortran and C should be able to be interoperable.

- The binding label with the BIND attribute and the other processing Fortran procedure names must not be same name.

- 308 -

- When C program is called as a function reference, the function must correspond as shown in "11.6 Passing and Receiving FunctionValues".

- When a Fortran program is called from a C program, the Fortran program must not be a main program.

- If the main program is a C program, the name must be "MAIN__". However, if the compiler option -mlcmain=main is specified, maincan be used for the function name.

- When using an interrupt-associated Fortran function, do not use signal(2) or clock(3C).

- When signal(2) or clock(3C) function is used in C language programs, the -i option must be specified at execution time to suppressinterrupt processing by Fortran objects and followings are restrictions.

- Specifying time limit to continue execution by -t option at execution time

- Service function "ALARM"

- Service function "SIGNAL"

- Always close nonstandard files before calling the other language.

- A CALL statement with an alternate return is valid only when calling Fortran subroutines.

- A subroutine called from a C program must not contain a RETURN statement with an integer expression or an aster of dummyargument. Change such a subroutine to an integer function.

- Fortran procedure must not invokes setjmp or longjmp in C program.

- If C program invokes setjmp or longjmp, that procedure must not cause any Fortran procedure to be invoked.

- When a Fortran procedure is called as a signal processing procedure, the signal processing procedure can refer only the variable thathave VOLATILE attribute.

- Fortran procedures are not allowed to call C language function with a variable number of arguments.

- If Fortran variable associates with C external variable (see Section "11.8 Passing Using External Variables") and the C external variablehas an initial value, the boundary of Fortran variable is changed to the boundary of the C external variable then warning message isoutput by linker.

- The variable of the language entity C_PTR type of the intrinsic module ISO_C_BINDING cannot debug using the debugger.

- If C++ program invokes exception processing (try, catch or throw), that procedure must not cause any Fortran procedure to be invoked.

- See "Notes on Interlanguage Linkage" in the "C++ User's Guide" for how to link C++ program.

- 309 -

Chapter 12 MultiprocessingThis chapter describes how parallel processing is performed for a Fortran program in this processor.

12.1 Overview of MultiprocessingThis section outlines the features of parallel processing and parallelization for Fortran in this system.

12.1.1 Parallel processingAlthough parallel processing has a broad meaning, the parallel processing mentioned in this section means to perform a single programsimultaneously using multiple CPUs working independently. In this section, parallel processing does not mean "multi job", which meansrunning multiple programs at the same time.

The following figure illustrates parallel processing using multiple CPUs simultaneously.

Figure 12.1 Multiprocessing

DO I = 1,50 A(I) = B(I) + C(I)END DO

Different iterations of the DO loop are executed ondifferent CPUs simultaneously.

DO I1 = 1,25 A(I1) = B(I1) + C(I1)END DO

DO I2 = 26,50 A(I2) = B(I2) + C(I2)END DO

CPU0 CPU1

12.1.2 Effect of MultiprocessingThe aim of parallel processing is to reduce the elapsed execution time of a program using multiple CPUs at the same time. For example,as shown in "Figure 12.1 Multiprocessing", ideally the elapsed time of the program is halved by splitting one DO loop into two andperforming them simultaneously ("Figure 12.2 Reducing elapsed time by using multiprocessing").

Figure 12.2 Reducing elapsed time by using multiprocessing

Elapsed

Time

CPU0 CPU0 CPU1

DO I=1,25 DO I=1,25 DO I=26,50

DO I=26,50 The elapsed time is reduced by executing on two CPUs.

Serial processing on one CPU Multiprocessing on two CPUs

- 310 -

Although the elapsed time of a program can be reduced using parallel processing, the CPU time of the program may not be reduced. Thisis because, even though a number of CPUs are used in parallel processing, the total CPU time used may be the same as the CPU time usedin serial processing, or may be increased because of the overhead to perform parallel processing.

However, in general, the performance of a program is evaluated by the elapsed time rather than the CPU time. If a program first executedserially is then executed in parallel, the total CPU time may be longer because of the overhead required for parallel processing.

12.1.3 Requirements for Effective MultiprocessingA computer environment that can use multiple CPUs at one time is required in order for parallel processing to reduce elapsed time. Althoughprograms generated by this compiler can run on hardware with a single CPU, if this is the case, a reduced elapsed time cannot be expected.Even if the hardware has multiple CPUs, a reduced elapsed time may not be expected when the CPU time is consumed by other jobs. Thisis because the opportunity to assign multiple CPUs at the same time to a program performing parallel processing may decrease.

In other words, for maximum benefit, parallel processing must be executed in an environment with multiple CPUs and sufficient capacityto assign multiple CPUs to the program.

Further, to minimize the relative rate of overhead due to parallel processing, the program should have either a large number of loopiterations or a large number of instructions within a loop.

12.1.4 Features of Parallel Processing in this CompilerThis compiler provides automatic parallelization and parallelization based on the OpenMP specifications.

The automatic parallelization is a feature of the compiler to parallelize the program automatically. Automatic parallelization is appliedonly for statements in DO loops and array operations for which the compiler has recognized that parallelization can be applied, but it doesnot require additional development.

This compiler also supports optimization control lines, which promote automatic parallelization. Information received from optimizationcontrol lines can be used to generate efficient object code for automatic parallelization.

To apply automatic parallelization, it is not necessary to update source code that has been designed for serial processing. For this reason,the source code is easily reusable in another system, as long as it conforms to the standard FORTRAN requirements.See Section "12.2Automatic Parallelization" for information on automatic parallelization.

The compiler also allows OpenMP parallelization. See Section "12.3 Parallelization by OpenMP Specifications" for information onparallelization in the OpenMP specification.

12.2 Automatic ParallelizationThis section explains the automatic parallelization provided by this compiler.

12.2.1 Compilation and ExecutionThe section explains how to compile and run a program that uses automatic parallelization.

12.2.1.1 CompilationTo activate automatic parallelization, specify the compiler option -Kparallel.

12.2.1.1.1 Compiler Option for Automatic Parallelization

The meaning and format of the automatic parallelization options are described below. See Section "2.2 Compiler Options" for moreinformation on each option.

-Kparallel[ ,reduction,instance=N,array_private,parallel_iteration=N,dynamic_iteration,ocl,optmsg=2,independent=proc_name,loop_part_parallel,region_extension ]

parallel

Automatic parallelization is performed. It automatically activates the -O2 and -Kregion_extension option. If the -H or -Eg option isset, the -Kparallel option will be ignored.

- 311 -

reduction

This option is effective when it is specified with the -Kparallel option. The reduction optimization (refer to Section "12.2.3.1.8 LoopReduction") is performed. This optimization may introduce differences in precision. Diagnostic messages output at compilation identifywhether or not this optimization was performed.

The reduction optimization is also performed when -Kparallel and -Keval have been specified at the same time.

instance=N

This option is effective when it is specified with the -Kparallel option. The compiler generates an object program assuming that theruntime thread number is N. Specifying this option allows the compiler to skip calculating the iteration count of DO loops, improvingthe performance of the compilation.

array_private

This option is effective when it is specified with the -Kparallel option. It privatizes arrays in a loop that can be privatized.

parallel_iteration=N 1 <= N <= 2147483647

This option is effective when it is specified with the -Kparallel option. It directs that only loops determined at compilation to have aniteration count exceeding N will be parallelized.

dynamic_iteration

This option is effective when it is specified with the -Kparallel option. It directs to dynamically select a loop to be parallelizedconsidering the number of threads, when both the inner loop and the outer loop within a nested loop are candidates.

ocl

This option is effective when it is specified with the -Kparallel option. It directs to activate the optimization control lines that are usedfor the automatic parallelization (see Section "12.2.3.2 Optimization Control Line").

optmsg=2

Messages are output that indicate that optimizations, such as automatic parallelization, SIMD optimization, and loop unrolling, havebeen performed.

independent=proc_name

This option is effective when it is specified with the -Kparallel option. It directs that the procedure specified in the pgm_nm argumentalways performs as a serial process, even if the procedure reference is specified within a parallelized DO loop. This means that DO-loops with procedure references will be subject to automatic parallelization.

loop_part_parallel

This option is effective when it is specified with the -Kloop_fission option and the -Kparallel option. When a loop contains statementsto which parallelization can be applied and statements to which parallelization cannot be applied, the loop is divided into two or moreloops and automatic parallelization is applied to one of the loops. This optimization is applied to the innermost loop.

region_extension

Parallelization overhead is reduced by expanding parallelization.

This option is automatically set when the -Kparallel option is specified.

12.2.1.2 Execution ProcessWhen running a program compiled with automatic parallelization, the number of threads can be specified using the environment variablePARALLEL or OMP_NUM_THREADS. The stack area size for each thread can be specified using the environment variableTHREAD_STACK_SIZE or OMP_STACKSIZE. It is also possible to specify a synchronization scenario using the environment variableFLIB_SPINWAIT or OMP_WAIT_POLICY.

The other procedures are the same as those used in serial processing.

12.2.1.2.1 Environment Variable PARALLEL

The environment variable PARALLEL specifies the number of threads performed at the same time. See Section "12.2.1.2.2 Number ofThreads" for details on determining the number of threads.

- 312 -

12.2.1.2.2 Number of Threads

The priority of the factors that determine the number of threads executed concurrently is as follows.

1. The value specified in the environment variable PARALLEL (Note)

2. The value specified in the environment variable OMP_NUM_THREADS

3. Number of CPUs that can be used with jobs

4. 1 thread

(Note)

If the -Kopenmp option has been specified at the linkage phase of the compilation and the environment variable FLIB_FASTOMPis TRUE, the value of the environment variable PARALLEL must be the same as the value of the environment variableOMP_NUM_THREADS. When the value does not match, a message will be output, and smaller value will be used for the numberof threads.

The number determined above is compared with the available maximum CPU number, and smaller number is used as the number ofthreads.

Refer to Section "D.1 Management of the CPUs resource" for information on the number of CPUs that can be used with jobs.

The maximum number of CPUs is determined in the following way.

- For jobs:

Number of CPUs that can be used with jobs

- Not for job:

Number of CPUs in the system

See Section "12.3.1.2.2 Environment Variable for OpenMP Specifications" for information on the environment variableOMP_NUM_THREADS.

12.2.1.2.3 Environment variables THREAD_STACK_SIZE and OMP_STACKSIZETHREAD_STACK_SIZE

Specify the stack size (in Kbytes) for each thread using the environment variable THREAD_STACK_SIZE.

When the environment variable OMP_STACKSIZE is also specified, the greater value of the two is used as the stack area size foreach thread. See Section "12.2.1.2.4 Stack Size on Execution" for information on stack area.

OMP_STACKSIZE

The stack area size for each thread can be specified using the environment variable OMP_STACKSIZE in byte, Kbytes, Mbytes, orGbytes.

When the environment variable THREAD_STACK_SIZE is also specified, the greater value of the two is used as the stack area size.See Section "12.2.1.2.4 Stack Size on Execution" for information on stack area.

12.2.1.2.4 Stack Size on Execution

Local scalar variables within a parallelized DO loop (except the automatic data object when the compiler option -Knoautoobjstack is ineffect) are allocated in the stack area for each thread. Allocate sufficient stack area, when there are many variables of this type.

The stack size for each thread is the same as the stack size for the process. However, if the maximum stack size for the process is largerthan the value calculated by the following expression:

min(the size of available memory in this system, the size of a virtual memory) / (the number of threads)

the system will determine the stack area size and allocate it as follows:

However, the assumed value does not guarantee that the system works correctly. It is recommended that an appropriate value for maximumprocess stack size be specified manually. The maximum process stack area size can be specified using the bash built in command "ulimit".Note that the maximum virtual memory size can also be confirmed using ulimit (bash built-in command).

S = (M/T)/5

- 313 -

S: the stack size for each thread (byte)M: the smaller size of available memory in this system and a virtual memory (byte)T: the number of threads

See Section "12.2.1.2.2 Number of Threads" for details. When the stack area size for each thread needs to be specified, use the environmentvariable THREAD_STACK_SIZE or OMP_STACKSIZE.

12.2.1.2.5 The Process of Waiting Threads

The standby threads in a serial process can be controlled with the environment variable FLIB_SPINWAIT or OMP_WAIT_POLICY.

FLIB_SPINWAIT value:

unlimited : Uses spin wait.

0 : Uses suspending wait.

<n>s : Uses spin wait until after <n> seconds, and uses suspending wait.

<n>[ms] : Uses spin wait until after <n> milliseconds, and uses suspending wait.

<n> is integer value that must be positive.

The default is unlimited.

Spin waiting is a technique to wait while consuming CPU time. Suspend waiting is a technique to wait without consuming CPU time.Since spin wait is more efficient in parallel execution, "unlimited" is suitable for reducing elapsed time. On the other hand, "0" issuitable for reducing CPU time.

OMP_WAIT_POLICY value:

ACTIVE : Uses spin wait.

PASSIVE : Uses suspending wait.

The default is "ACTIVE".

Select ACTIVE to give priority to the elapsed time. Select PASSIVE to give priority to the CPU time.

The settings of environment variable FLIB_SPINWAIT will be valid when environment variable FLIB_SPINWAIT is specified.

12.2.1.2.6 Notes when setting time limits

Operation may not be correct when time limits are set with the limit(1) command. If operation is not correct, use the Fortran exceptionhandling process, the runtime option -Wl, -i to invalidate it. Alternatively, use the runtime option -Wl,-t to set the time limit.

12.2.1.2.7 Notes when using service routines

- ALARM Service FunctionIt may not work correctly. It is only for a program using serial processing.

- KILL and SIGNAL Service FunctionsIt may cause a deadlock. It is only for a program using serial processing.

- FORK Service FunctionIt may not work correctly. It is only for a program using serial processing.

- SYSTEM, SH and CHMOD Service FunctionsMemory usage doubles and performance may deteriorate. Use only with programs using serial processing.

12.2.1.2.8 CPU Binding for Thread

CPU Binding for Thread differs between job execution and non-job execution.

When executing program on non-job.

The thread is not bound to a CPU. But you can control CPU Binding for Thread using environment variable FLIB_CPU_AFFINITY.

- 314 -

The environment variable FLIB_CPU_AFFINITY

Threads are bound to CPUs in order of the specified cupid list.

When the number of specified CPUs is exceeded, it is repeatedly used from the beginning of the list.

The cpuid shall be separated by comma (',') or space(' ').

The cpuid list can have the next form that has range with increment.

cpuid1[-cpuid2[:inc]]

cpuid1: cpuid of the beginning of the range. (0<=cpuid1<CPU_SETSIZE)

cpuid2: cpuid of the end of the range. (0<=cpuid2<CPU_SETSIZE)

inc: increment (1<=inc<CPU_SETSIZE)

In addition, it is necessary to be the following.

cpuid1<=cpuid2

It becomes equivalent to the case where all CPUs for every increment value inc in the range from cpuid1 to cpuid2 are specified.

The cpuid can be used the above-mentioned value. However, cpuid which can actually be used becomes only within the limits of CPUaffinity of the process at the time of an execution start.

See CPU_SET(3) about details of CPU_SETSIZE.

Examples of using environment variable FLIB_CPU_AFFINITY are shown below.

Example 1:

$ export FLIB_CPU_AFFINITY="0,2,1,3"

The thread is bound to CPU in order of 0, 2, 1, and 3.When the number of threads is five or more, it is repeatedly used from the beginning of the list.

Example 2:

$ export FLIB_CPU_AFFINITY="0-7"

The thread is bound to CPU in order of 0, 1, 2, 3, 4, 5, 6 and 7.When the number of threads is nine or more, it is repeatedly used from the beginning of the list.

Example 3:

$ export FLIB_CPU_AFFINITY="0-7:2"

The thread is bound to CPU in order of 0, 2, 4 and 6.When the number of threads is five or more, it is repeatedly used from the beginning of the list.

Example 4:

$ export FLIB_CPU_AFFINITY="0-4:2,1,7"

The thread is bound to CPU in order of 0, 2, 4, 1 and 7.When the number of threads is six or more, it is repeatedly used f from the beginning of the list.

When executing program on job.

When thread is not bound to CPU, you may control CPU Binding with FLIB_CPU_AFFINITY as executing on non-job.

When thread is bound to CPU on job, the environment variable FLIB_CPU_AFFINITY becomes invalid.

For details of CPU Binding, see Section "D.2 CPU Binding".

- 315 -

12.2.1.3 Example of Compilation and ExecutionExample 1:

$ frtpx -Kparallel,reduction,ocl test1.f$ ./a.out

Example 1 compiles the source program using the reduction optimization while enabling optimization control lines. This programperforms using a single thread.

Example 2:

$ frtpx -Kparallel -Koptmsg=2 test2.fFortran diagnostic messages: program name (main) jwd5001p-i "test2.f", line 2: DO loop with DO variable 'i' is parallelized.$ export PARALLEL=2$ ./a.out$ export PARALLEL=4$ ./a.out

Example 2 compiles specifying the -Koptmsg=2 option to confirm that automatic parallelization is performed. By specifying 2 for theenvironment variable PARALLEL, the program executes with two threads. The next step executes the program with four threads, byspecifying 4 for PARALLEL.

12.2.2 Tuning of parallelized programsA Sampling feature can be used to improve the performance of a parallelized program. The sampling feature helps to identify the statementsin a program that are expensive. Designing a program to perform expensive statements simultaneously can improve its performance. Onthe Job Operation Software environment, use the sampling feature provided by the Programming Workbench to monitor the executiontime.

For other environments, use the sampling feature provided by the profiler to monitor the execution time. To collect information about theparallel program using the sampling feature, specify an environment variable on executing the program.

See the "Programming Workbench User's Guide" for information on the programming workbench. See the "Profiler User's Guide" forinformation on the Profiler.

12.2.3 Details of Automatic ParallelizationThis section describes optimization control lines, which are used to implement automatic parallelization. It also describes the points thepoints to note when using automatic parallelization.

12.2.3.1 Automatic ParallelizationThis section describes the automatic parallelization.

12.2.3.1.1 Targets for Automatic Parallelization

Automatic parallelization is performed for DO loops (including nested DO loops) and array statements (array operation, array assignments)in the Fortran source code.

12.2.3.1.2 What is Loop Slicing?

Automatic parallelization may slice a loop into several pieces. The elapsed execution time is reduced by executing the loop slices inparallel. The following figure illustrates the image of the loop slicing.

Figure 12.3 Loop slicing

DO I=1,10000 A(I)=A(I)+B(I)END DO

- 316 -

DO I=1,5000 A(I)=A(I)+B(I)END DO

DO I=5001,10000 A(I)=A(I)+B(I)END DO

12.2.3.1.3 Array Operations and Automatic Parallelization

In addition to DO loops, automatic parallelization is also performed on array statements (array operation, array assignments.) The followingfigure illustrates automatic parallelization on array operation.

Figure 12.4 Automatic parallelization of a statement with an array operation

INTEGER A(1000),B(1000)A = A + B

A(1:500)=A(1:500)+B(1:500) A(501:1000)=A(501:1000)+B(501:1000)

12.2.3.1.4 Automatic Loop Slicing by the Compiler

The system parallelizes a loop if the order of data reference will be the same as with serial execution, and if the system is certain that theresult of the parallel loop will be the same as if the loop were processed serially.

"Figure 12.5 A loop that is not a candidate for loop slicing" shows an example where automatic loop slicing is not applied. In this DOloop, the value of the array element A(5000), which is defined when the loop counter I is 5000, is referenced when I is 5001. If loop slicingis performed for this DO loop, however, the element A(5000) may be referenced by the other loop before its value is defined, causing anerror.

"Figure 12.5 A loop that is not a candidate for loop slicing" Sample DO loop that is not supported by loop slicing

Figure 12.5 A loop that is not a candidate for loop slicing

DO I=2,10000 A(I)=A(I-1)+B(I)END DO

DO I=2,5000 A(I)=A(I-1)+B(I)END DO

DO I=5001,10000 A(I)=A(I-1)+B(I)END DO

12.2.3.1.5 Loop Interchange and Automatic Loop Slicing

When nested loops are sliced, the system attempts to parallelize the outermost loop if it can. Therefore, the system selects the loop thatcan be sliced most cheaply and interchanges it with the outermost loop. The purpose of this is to reduce the overhead of multiprocessingand improve the execution performance.

"Figure 12.6 Loop interchange in nested DO loops" illustrates the loop slicing after performing loop interchange on a nested DO loop.Before performing parallelization for the DO loop variable "J", loop interchange is performed so that the parallelization overhead isreduced.

- 317 -

Figure 12.6 Loop interchange in nested DO loops

DO I=2,10000 DO J=1,10 A(I,J)=A(I-1,J)+B(I,J) END DOEND DO

DO J=1,10 <-- Interchange DO I=2,10000 <-- A(I,J)=A(I-1,J)+B(I,J) END DOEND DO

DO J=1,5 DO I=2,10000 A(I,J)=A(I-1,J)+B(I,J) END DOEND DO

DO J=6,10 DO I=2,10000 A(I,J)=A(I-1,J)+B(I,J) END DOEND DO

12.2.3.1.6 Loop Distribution and Automatic Loop Slicing

In the DO loop in "Figure 12.7 Separation of DO loop and loop slicing", loop slicing cannot be performed for array A because the definitionor order of data references would be different from the order of data references in serial execution. However, loop slicing can be performedfor array B, because the order of parallel data references is the same as the order of serial data references. In this case, the statement wherearray A is defined and the statement where array B is defined are separated into two loops, and loop slicing is performed on the loop wherearray B is defined.

The partial automatic parallelization by the distribution of loop is performed when the -Kloop_part_parallel option is effective and thecompiler assumes that the distribution of loop is effective.

The following figure illustrates separating the statements into DO loops and performing loop slicing.

Figure 12.7 Separation of DO loop and loop slicing

DO I=2,10000 A(I)=A(I-1)+C(I) B(I)=B(I)+C(I)END DO

DO I=2,10000 A(I)=A(I-1)+C(I) <-- DistributionEND DO

DO I=2,10000 B(I)=B(I)+C(I) <-- DistributionEND DO

DO I=2,5000 B(I)=B(I)+C(I)END DO

DO I=5001,10000 B(I)=B(I)+C(I)END DO

- 318 -

12.2.3.1.7 Loop Fusion and Automatic Loop Slicing

"Figure 12.8 Fusion of DO loops and loop slicing" contains two DO loops with the same iteration count. In this case, combining the loopsinto one can reduce the overhead of both parallelization and loop control.

The following figure illustrates the loop slicing after loop fusion.

Figure 12.8 Fusion of DO loops and loop slicing

DO I=1,10000 A(I)=B(I)*C(I)END DO

DO I=1,10000 D(I)=E(I)+F(I)END DO

DO I=1,10000 A(I)=B(I)*C(I) <-- Fusion D(I)=E(I)+F(I) <--END DO

DO I=1,5000 A(I)=B(I)*C(I) D(I)=E(I)+F(I)END DO

DO I=5001,10000 A(I)=B(I)*C(I) D(I)=E(I)+F(I)END DO

12.2.3.1.8 Loop Reduction

If both the compiler options -Kparallel and -Kreduction are specified at the same time, the automatic parallelization loop reduction willbe performed. This performs loop slicing by changing the order of operations such as addition and multiplication if they are interchangeable.However, this may affect precision.

Loop reduction is performed only when the DO loop includes the following operations:

- SUM S=S+A(I)

- PRODUCT P=P*A(I)

- DOT PRODUCT P=P+A(I)*B(I)

- MIN X=MIN(X,A(I))

- MAX Y=MAX(Y,A(I))

- OR N=N.OR. A(I)

- AND M=M.AND.A(I)

- 319 -

The following figure shows an example of loop reduction and loop slicing.

Figure 12.9 Loop reduction

SUM=0DO I=1,10000 SUM=SUM+A(I)END DO

SUM1=0DO I=1,5000 SUM1=SUM1+A(I)END DO


SUM=SUM+SUM1+SUM2

12.2.3.1.9 Restrictions on Loop Slicing

Loop slicing does not support the following DO loops:

- If it is unlikely that parallelization will reduce elapsed time.

- If operation types not supported by loop slicing are included

- If references to external procedures, internal procedures, or module procedure are included When the form of the DO-loop is complex

- Input-output statements or intrinsic subroutines are included

- Data definition or order of data references order may differ from serial execution

- If it is unlikely that parallelization will reduce elapsed time

1. It can be forecast that the elapsed time would not be reduced in the case of loop slicing.

When the iteration counts of a DO loop is small or there are few statements within the DO loop, the parallelization overheadintroduced by loop slicing may deteriorate performance. The compiler will not perform loop slicing when performance improvementcannot be expected for the DO loop.

The following figure shows an example of a DO loop with a small iteration count and few statements.

Figure 12.10 Small loop iteration count and small number of operations

DO I=1,10 ! Not parallelized A(I)=A(I)+B(I) ! because of small loop iteration count and ! small number of operationsEND DO

2. The loop contains operations on types, which are not supported in loop slicing.

Loop slicing does not support DO loops or inner DO loop with the following types of operations:

- Character type

- Derived type

- 320 -

3. References to external procedures, internal procedures, or module procedure are included.

Loop slicing does not support DO loops containing a reference to an external procedure, internal procedure, or module procedure.It is also unsupported when an inner DO loop includes the following operation types.

However, optimization control lines may be used to promote parallelization. See Section "12.2.3.2 Optimization Control Line"fordetails.

Inline expanded external procedure and internal procedures are subject to parallelization.

The following figure shows an example of a DO loop that contains a subroutine call.

Figure 12.11 DO loop contains a subroutine call

DO J=1,10 ! Not parallelized DO I=1,10000 ! because of the subroutine call A(I,J)=A(I,J)+B(I,J) CALL SUB(A) END DOEND DO

4. DO loop with complicated structure

The following DO loops cannot be sliced because they are too complicated:

- There is a branch from inside the loop to outside the DO loop.

- The structure of the DO loop is complicated.

The following figure shows an example of a DO loop that has a jump from inside to outside the loop.

Figure 12.12 Jumping out from inside of the DO loop

DO J=1,10 ! Not parallelized DO I=1,10000 ! because of jumping out of the loop A(I,J)=A(I,J)+B(I,J) IF(A(I,J).LT.0) GO TO 20 END DO END DO . . . 20 CONTINUE

The following figure shows an example of DO loop with a complicated structure.

Figure 12.13 Complicated DO loop structure

DO J=1,10000 ! Not parallelized, because IF (N >0) THEN ! of complicated DO loop structure ASSIGN 10 TO I ELSE IF (N <0) THEN ASSIGN 20 TO I ELSE ASSIGN 30 TO I END IF GO TO I,(10,20,30) 10 A(J)=A(J)+0

- 321 -

20 A(J)=A(J)+1 30 A(J)=A(J)+2 END DO

5. Input-output statements or intrinsic subroutines are included.

Loop slicing does not support DO loops containing an input-output statement or intrinsic subroutine.

Intrinsic functions, which are not supported by loop slicing, can be confirmed using diagnostic messages.

6. Data definition or reference order may differ from the serial process

As shown in "Figure 12.5 A loop that is not a candidate for loop slicing", loop slicing does not support DO loops that change datadefinition or the order of data references when converted from serial execution.

12.2.3.1.10 Displaying the State of Automatic Parallelization

To confirm whether automatic parallelization is performed, specify compiler options -Kparallel and -Koptmsg=2 at the same time. Thecompilation diagnostic message will indicate the status.

When automatic parallelization is not performed, the message also outputs the reason.

The following shows an example of diagnosis message:

Example 1: Message when the program in "Figure 12.3 Loop slicing" is compiled

jwd5001p-i "test.f", line 2: DO loop with DO variable 'I' is parallelized.

Example 2: Message when the program in "Figure 12.5 A loop that is not a candidate for loop slicing" is compiled

jwd5201p-i "test.f", line 2: DO loop is not parallelized: data dependency array 'A' may cause different results from serial execution for loop.

12.2.3.1.11 Extension of Parallel Region

The Parallel region is the range between the creation of multiple threads for automatic parallelization, and the deallocation of each threadwhen each thread ends. In general, parallelization generates one parallelized loop for one parallel region as shown in "Figure 12.14 Exampleof extension of parallel region". The cost of generating a loop for parallelization during runtime is considered the parallelization overhead.

By expanding the parallel region for these two parallelization loops, two parallelized loops will be generated for one parallel region. Asa result, parallelization overhead is reduced, because the parallel region is generated only once.

Figure 12.14 Example of extension of parallel region

DO I=1, 1000 A(I) = B(I) + C(I)END DO

parallel loop parallel region

DO J=1, 2000 D(J) = E(J) + F(J)END DO


extension of parallel region

DO I=1, 1000 A(I) = B(I) + C(I)END DO

parallel loop

parallel regionDO J=1, 2000 D(J) = E(J) + F(J)END DO

parallel loop

- 322 -

The parallelization overhead can also be reduced for nested loops as shown in "Figure 12.15 Example of extension of parallel region(nested loop)" by expanding the parallel region. In this example, optimization reduces the number of times the parallel region is generatedfrom 1000 down to 1.

Figure 12.15 Example of extension of parallel region (nested loop)

DO I=1, 1000 DO J=1, 2000 A(J) = B(J) + C(J) END DOEND DO


extension of parallel region

DO I=1, 1000 DO J=1, 2000 A(J) = B(J) + C(J) END DOEND DO


12.2.3.1.12 Block Distribution and Cyclic Distribution

The methods of loop slicing in the automatic parallelization are block distribution and cyclic distribution.

The block distribution is to allocate the block that the loop iteration count is distributed by the number of threads.

The cyclic distribution is to sequentially allocate the block that the loop iteration count is distributed by the arbitrary size. The defaultmethod of the automatic parallelization is block distribution. When PARALLEL_CYCLIC optimization control specifier is specified, theDO loop is cyclically distributed.

Figure 12.16 Block distribution and cyclic distribution

SUBROUTINE SUB(A,B)INTEGER A(20,20),B(20,20)DO J=1,20 ! parallelized DO I=J,20 A(I,J) = B(I,J) ENDDOENDDOEND

When block distribution or cyclic distribution is applied to "Figure 12.16 Block distribution and cyclic distribution", the loop iterationcount is allocated to each thread as follows:

- Block distribution

The block that the loop iteration count is divided by number of threads (2) is allocated in each thread as follows:

Thread 0: {1,2,3,4,5,6,7,8,9,10}Thread 1: {11,12,13,14,15,16,17,18,19,20}

- Cyclic distribution (block size: 2)

The block that the loop iteration count is divided by two iteration (block size: 2) is sequentially allocated in each thread as follows:

- 323 -

Thread 0: {1,2}, {5,6}, { 9,10}, {13,14}, {17,18}Thread 1: {3,4}, {7,8}, {11,12}, {15,16}, {19,20}

12.2.3.2 Optimization Control LineThis system provides OCLs (Optimization Control Lines) to control automatic parallelization.

To activate the OCL, specify the compiler options -Kparallel and -Kocl at the same time.

12.2.3.2.1 Notation rules for optimization control lines

The column 1 to 5 of the OCL must be "!OCL ". One or more optimization specifiers for automatic parallelization follows the "!OCL ".The notation rules are detailed in the following sections.

!OCL i [,i] ....

i: Optimization control specifier for automatic parallelization

12.2.3.2.2 Description position of the optimization control line

The position where the optimization control line can be inserted depends on the type of optimization indicator.

The OCL for automatic parallelization must be specified at the position of total or loop. These are defined as follows:

- Total-positionAt the beginning of each program unit.

- Loop-positionImmediately in front of DO statements. Other OCLs that can be specified for loop position or comment lines can be specified betweenthe OCL and DO statement.

Example:

!OCL SERIAL <------------------ total-position SUBROUTINE SUB(B,C,N) INTEGER A(N),B(N),C(N) DO I=1,N A(I)=B(I)+C(I) END DO PRINT *, FUN(A)!OCL PARALLEL <---------------- loop-position DO I=1,N A(I)=B(I)*C(I) END DO PRINT *, FUN(A) END SUBROUTINE

12.2.3.2.3 Automatic Parallelization and Optimization Control Specifiers

The Optimization Specifiers for automatic parallelization are ignored if they are specified for a DO loop for which loop slicing is notsupported. See Section "12.2.3.1.9 Restrictions on Loop Slicing"for information about DO loops which are not supported by loop slicing.

12.2.3.2.4 Optimization Control Specifiers for Automatic Parallelization

The functionality of OCLs differs depending on the optimization specifier.

The optimization control specifiers for automatic parallelization specify useful information for the compiler to generate an efficient objectmodule. They are listed in the table below.

Table 12.1 List of optimization control specifiers for automatic parallelization


ARRAY_PRIVATE Designates to be privatized arrays.

NOARRAY_PRIVATE Designates canceling to be privatized arrays.

- 324 -


INDEPENDENT [(e[,e]...)] Asserts that the procedure (e) has no side effects, and that the DO loop that contains areference to the procedure (e) can be sliced.

LOOP_PART_PARALLEL Designates automatic parallelization that requires dividing loops.

LOOP_NOPART_PARALLEL Suppresses automatic parallelization that requires dividing loops.

SERIALPARALLEL

Controls whether the compiler will attempt automatic parallelization or not.

PARALLEL_CYCLIC[(n)] Designates cyclic distribution in block size (n).

PARALLEL_STRONG Designates to parallelize the loop that the loop has small number of loop iterations orthe loop has small the number of operations.

REDUCTION Designates to parallel reduction operations.

NOREDUCTION Designates canceling to parallel reduction operations.

TEMP [(var[,var]...)] Indicate variables (var) that are only used in the following DO loop and not outside ofthe loop

TEMP_PRIVATE(var[,var]...)Designates to assign variable (var) to local area in thread. As a result, the datadependency between threads is solved, and the automatic parallelization is promoted.

FIRST_PRIVATE(var[,var]...)Designates to assign variable (var) to local area in thread. And, designates to have apossibility of reference of the initial value of variable. As a result, the data dependencybetween threads is solved, and the automatic parallelization is promoted.

LAST_PRIVATE(var[,var]...)Designates to assign variable (var) to local area in thread. And, the variable is directedto be referred after this DO loop. As a result, the data dependency between threads issolved, and the automatic parallelization is promoted.

var1 op var2 or var1 op cnst Designates relationship between variable and variable or between variable andconstant.

The details of the optimization control specifiers are described below.

For clarity, the source code used in this section has; (p) for statements that are parallelized, (m) for statements that are partially parallelized,and (s) for statements that are not parallelized.

ARRAY_PRIVATE

The ARRAY_PRIVATE specifier directs the system to identify arrays in a loop that can be privatized, and promotes parallelizationby privatizing them.

The ARRAY_PRIVATE indicator has the following format:

!OCL ARRAY_PRIVATE

The ARRAY_PRIVATE indicator can be entered in the loop position and total position.

Its effect differs depending on the position.

- In the loop position

it is effective for the corresponding DO loop.

- In the total position

It is effective for all DO loops in the corresponding program unit.

In "Figure 12.17 Usage of the ARRAY_PRIVATE specifier", using the ARRAY_PRIVATE specifier, the compiler detects that arrayA can be privatized and it performs parallelization by privatizing the arrays, even if the compiler option -Karray_private is not specified.

Figure 12.17 Usage of the ARRAY_PRIVATE specifier

- 325 -

SUBROUTINE SUB(N,M) INTEGER(KIND=4),DIMENSION(10000):: A INTEGER(KIND=4),DIMENSION(10000,100) :: B,C,D DATA A/10000*1/, C/1000000*2/, D/1000000*3/ !OCL ARRAY_PRIVATE p DO I=1,N p DO J=1,M p A(J) = I + D(J,I) p B(J,I) = A(J) + C(J,I) p ENDDO p ENDDO PRINT*, A,B END

NOARRAY_PRIVATE

The NOARRAY_PRIVATE specifier directs the compiler not to privatize arrays even if they can be privatized.

The NOARRAY_PRIVATE indicator has the following format:

!OCL NOARRAY_PRIVATE

The NOARRAY_PRIVATE indicator can be entered in the loop position and total position. Its effect differs depending on the position.

- At the loop-position

It is effective for the corresponding DO loop.

- At the total-position

It is effective for all the DO loops in the corresponding program units.

In "Figure 12.18 Usage of the NOARRAY_PRIVATE specifier", array A is not privatized due to NOARRAY_PRIVATE, althoughit can be privatized and the compiler option -Karray_private is set.

Figure 12.18 Usage of the NOARRAY_PRIVATE specifier

SUBROUTINE SUB(N,M) INTEGER(KIND=4),DIMENSION(10000):: A INTEGER(KIND=4),DIMENSION(10000,100) :: B,C,D DATA A/10000*1/, C/1000000*2/, D/1000000*3/ !OCL NOARRAY_PRIVATE p DO I=1,N p DO J=1,M p A(J) = I + D(J,I) p B(J,I) = A(J) + C(J,I) p ENDDO p ENDDO PRINT*, A,B END

INDEPENDENT[(e[,e]...)]

The INDEPENDENT specifier directs the compiler that performing the procedure call within the DO loop in parallel execution willnot change the operation result compared to serial execution. Doing this means that DO-loops with procedure references will be subjectto loop slicing.

The INDEPENDENT indicator has the following format:

!OCL INDEPENDENT[(e[,e]...)]

- 326 -

'e' are the procedure names that will not be affected by loop slicing. Wildcards can be specified in 'e'. When no procedure name isspecified, this specifier is applied to all of the procedures within the target range. See Section "12.2.3.2.5 Wild Card Specification" forinformation on specifying wildcards.

The procedures specified in 'e' must be compiled with the compiler options -Kthreadsafe and -Kparallel.

The INDEPENDENT indicator can be entered in the loop position and total position.


- At the loop-positionIt is effective for the corresponding DO loop.

- At the total-positionIt is effective for all the DO loops in the corresponding program units.

Figure 12.19 Example program without INDEPENDENT specifier

REAL FUN,A(10000)DO I=1,10000 J=I A(I)=FUN(J)END DO . .ENDFUNCTION FUN(J)INTEGER JREAL FUNFUN=SQRT(REAL(J**2+3*J+6))RETURNEND

In "Figure 12.19 Example program without INDEPENDENT specifier", the procedure FUN is called within the DO loop and thecompiler cannot identify whether loop slicing can be performed for this loop.

If performing loop slicing for the DO loop that calls the procedure FUN does not affect the result, using the INDEPENDENT specifieras in "Figure 12.20 Usage of INDEPENDENT specifier" will perform loop slicing for the DO loop.

Figure 12.20 Usage of INDEPENDENT specifier

REAL FUN,A(1000) !OCL INDEPENDENT(FUN) p DO I=1,1000 p J=I p A(I)=FUN(J) p END DO . . END RECURSIVE FUNCTION FUN(J) INTEGER J REAL FUN FUN=SQRT(REAL(J**2+3*J+6)) RETURN END

Note:

- 327 -

Ensure that INDEPENDENT is only specified for procedures for which loop slicing is supported. The following shows an exampleof a procedure type that is supported by loop slicing:

- Procedures that have dependency relationship with other procedures

LOOP_PART_PARALLEL

The LOOP_PART_PARALLEL specifier directs the system to perform automatic parallelization that requires dividing loops.

The LOOP_PART_PARALLEL indicator has the following format:

!OCL LOOP_PART_PARALLEL

The LOOP_PART_PARALLEL indicator can be entered in the loop position and total position.


- In the loop positionIt is effective for the corresponding DO loop.

- In the total positionIt is effective for all DO loops in the corresponding program unit.

The LOOP_PART_PARALLEL specifier is in effect for the code shown in "Figure 12.21 Usage of the LOOP_PART_PARALLELspecifier", then the loop is divided and automatic parallelization is applied to the part of the divided loop without the -Kloop_part_parallel option.

Figure 12.21 Usage of the LOOP_PART_PARALLEL specifier

REAL(KIND=8) A(N),B(N),C(N),D(N) !OCL LOOP_PART_PARALLELm DO I=2,N p A(I) = A(I)-B(I)+LOG(C(I)) ! automatic parallelization can be applied to ! this statements D(I) = D(I-1)+A(I) ! automatic parallelization cannot be applied to ! this statementp ENDDO

LOOP_NOPART_PARALLEL

The LOOP_NOPART_PARALLEL specifier directs to suppress automatic parallelization that requires dividing loops.

The LOOP_NOPART_PARALLEL indicator has the following format:

!OCL LOOP_NOPART_PARALLEL

The LOOP_NOPART_PARALLEL indicator can be entered in the loop position and total position.


- In the loop positionIt is effective for the corresponding DO loop.

- In the total positionIt is effective for all DO loops in the corresponding program unit.

The LOOP_NOPART_PARALLEL specifier is in effect for the code shown in "Figure 12.22 Usage of theLOOP_NOPART_PARALLEL specifier", then the loop is not divided and automatic parallelization is not performed to the loop.

Figure 12.22 Usage of the LOOP_NOPART_PARALLEL specifier

REAL(KIND=8) A(N),B(N),C(N),D(N) !OCL LOOP_NOPART_PARALLEL

- 328 -

s DO I=2,Np A(I) = A(I)-B(I)+LOG(C(I))s D(I) = D(I-1)+A(I)p ENDDO

SERIAL

The SERIAL indicator is used to prevent loop slicing of DO loops.

This specifier can be used for a DO loop that is faster when executed serially.

The SERIAL indicator has the following format:

!OCL SERIAL

The SERIAL indicator can be entered in the loop position and total position. Its effect differs depending on the position.

- At the loop-positionIt prevents loop slicing for the corresponding DO loop and its inner DO loops.

- At the total-positionIt prevents loop slicing for all the DO loops in the corresponding program units.

To suppress loop slicing for the sample program in "Figure 12.23 Example program without SERIAL specifier", position the SERIALspecifier as in "Figure 12.24 Usage of the SERIAL specifier".

Figure 12.23 Example program without SERIAL specifier

REAL,DIMENSION(100,100) :: A2,B2,C2,D2,E2,F2,G2,H2 p DO J=1,100 p DO I=1,M p A2(I,J)=A2(I,J)+B2(I,J) p C2(I,J)=C2(I,J)+D2(I,J) p E2(I,J)=E2(I,J)+F2(I,J) p G2(I,J)=G2(I,J)+H2(I,J) p END DO p END DO END

Figure 12.24 Usage of the SERIAL specifier

REAL,DIMENSION(100,100) :: A2,B2,C2,D2,E2,F2,G2,H2!OCL SERIAL DO J=1,100 DO I=1,M A2(I,J)=A2(I,J)+B2(I,J) C2(I,J)=C2(I,J)+D2(I,J) E2(I,J)=E2(I,J)+F2(I,J) G2(I,J)=G2(I,J)+H2(I,J) END DO END DO END

PARALLEL

The PARALLEL specifier is used to suppress the effect of the SERIAL specifier and perform loop slicing for specific DO loops.

- 329 -

PARALLEL indicator format:

!OCL PARALLEL

The PARALLEL indicator can be entered in the loop position and total position. Its effect differs depending on the position.

- At the loop-positionIt performs loop slicing for the corresponding DO loop and its inner DO loop.

- At the total-positionIt pertains to all the DO loops in the corresponding program unit.

To perform loop slicing for loop 2 in "Figure 12.25 Example program without PARALLEL specifier", specify the SERIAL andPARALLEL specifiers as in "Figure 12.26 Example program with combination of PARALLEL and SERIAL specifiers" so that loopslicing is performed only for the DO loop presented in 2.

Figure 12.25 Example program without PARALLEL specifier

SUBROUTINE SUB(A1,A2,A3,B1,B2,B3,M1,M2,M3,N1,N2,N3) INTEGER :: M1,M2,M3,N1,N2,N3 REAL,DIMENSION(M1,N1) :: A1,B1 REAL,DIMENSION(M2,N2) :: A2,B2 REAL,DIMENSION(M3,N3) :: A3,B3

p DO J=1,N1 ! loop 1p DO I=1,100p A1(I,J) = A1(I,J) + B1(I,J)p ENDDOp ENDDO



RETURN END

Figure 12.26 Example program with combination of PARALLEL and SERIAL specifiers

!OCL SERIAL SUBROUTINE SUB(A1,A2,A3,B1,B2,B3,M1,M2,M3,N1,N2,N3) INTEGER :: M1,M2,M3,N1,N2,N3 REAL,DIMENSION(M1,N1) :: A1,B1 REAL,DIMENSION(M2,N2) :: A2,B2 REAL,DIMENSION(M3,N3) :: A3,B3

DO J=1,N1 ! loop 1 DO I=1,100 A1(I,J) = A1(I,J) + B1(I,J) ENDDO ENDDO

- 330 -

!OCL PARALLELp DO J=1,N2 ! loop 2p DO I=1,100p A2(I,J) = A2(I,J) + B2(I,J)p ENDDOp ENDDO

DO J=1,N3 ! loop 3 DO I=1,100 A3(I,J) = A3(I,J) + B3(I,J) ENDDO ENDDO

RETURN END

PARALLEL_CYCLIC[(n)]

The PARALLEL_CYCLIC specifier directs the system to perform cyclic distribution for a DO loop that automatic parallelization canbe analyzed by compiler. When this specifier is applied, loop is parallelized without estimating the effects of the parallelization.

See Section "12.2.3.1.12 Block Distribution and Cyclic Distribution" for information on cyclic distribution.

The PARALLEL_CYCLIC indicator has the following format:

!OCL PARALLEL_CYCLIC[(n)]

n is the block size of cyclic distribution. n is number from 1 to 10000. The default is 1.

The PARALLEL_CYCLIC indicator can be entered in the loop position. It is effective for the corresponding DO loop at the loopposition.

In "Figure 12.27 Usage of the PARALLE_CYCLIC specifier", DO loop is cyclically distributed with block size of 2.

Figure 12.27 Usage of the PARALLE_CYCLIC specifier

SUBROUTINE SUB(A,B,N) INTEGER A(N,N),B(N,N) !OCL PARALLEL_CYCLIC(2) p DO J=1,N p DO I=J,N p A(I,J) = B(I,J) p ENDDO p ENDDO END

PARALLEL_STRONG

The PARALLEL_STRONG specifier directs to perform loop slicing for loops which have not been sliced because the iteration countor instruction numbers are too small.

The PARALLEL_STRONG indicator has the following format:

!OCL PARALLEL_STRONG

The PARALLEL_STRONG indicator can be entered in the loop position and total position. Its effect differs depending on the position.

- At the loop-positionIt pertains to the corresponding DO loop and its inner DO loops.

- 331 -


The PARALLEL_STRONG specifier forces parallelization for the loop in Figure 12.28 Usage of PARALLEL_STRONG specifier,although the loop is usually excluded from parallelization because the iteration count is considered too small to gain any benefit.

Figure 12.28 Usage of PARALLEL_STRONG specifier

REAL(KIND=4) :: A(10),B(10),C(10) !OCL PARALLEL_STRONG p DO J=1,10 p A(J) = A(J) + B(J) + C(J) p END DO PRINT*, A END

REDUCTION

The REDUCTION specifier directs the compiler to perform parallelization for a DO loop that includes reduction operations.

The REDUCTION indicator has the following format:

!OCL REDUCTION

The REDUCTION indicator can be entered in the loop position and total position. Its effect differs depending on the position.



As in "Figure 12.29 Usage of the REDUCTION specifier", the REDUCTION specifier forces parallelization for the DO loop thatincludes reduction operations, even if the compile option -Kreduction is not specified.

Figure 12.29 Usage of the REDUCTION specifier

REAL S,A(5000) !OCL REDUCTION p DO I=1,5000 p S = S + A(I) p END DO END

NOREDUCTION

The NOREDUCTION specifier directs the compiler not to perform parallelization for a DO loop that includes reduction operations.

The NOREDUCTION indicator has the following format:

!OCL NOREDUCTION

The NOREDUCTION specifier may be specified at the loop-position or the total-position. The effect of NOREDUCTION depend onits position.



- 332 -

As illustrated in "Figure 12.30 Usage of the NOREDUCTION specifier", the NOREDUCTION specifier can suppress parallelizationof the specified DO loop although it includes the reduction operation.

Figure 12.30 Usage of the NOREDUCTION specifier

REAL S,A(5000) !OCL NOREDUCTION s DO I=1,5000 s S = S + A(I) s END DO END

TEMP[(var[,var]...)]

The TEMP specifier directs the compiler that the variable used in the DO loop is a temporary variable, which is only valid within theDO loop. This can improve the performance of the parallelized DO loop.

The TEMP indicator has the following format:

!OCL TEMP[(var[,var]...)]

'var' is a temporary variable name used in the DO-loop. Wildcards can be specified in 'var'. When no variable name is specified, thisspecifier is applied to all of the variables within the target range. See Section "12.2.3.2.5 Wild Card Specification" for information onspecifying wildcards.

The TEMP indicator can be entered in the loop position and total position. Its effect differs depending on the position.

- At the loop-positionIt directs that the specified variables, which are used in the corresponding DO loop, are temporary variables.

- At the total-positionIt directs that the specified variables, which are used in all DO loops in the corresponding program unit, are temporary variables.

In "Figure 12.31 Example program without a TEMP specifier", the compiler considers that "T" may be used in the SUB because "T"is declared in the common block, and the compiler performs loop slicing while maintaining the value of "T" although it is only usedwithin the DO loop.

Figure 12.31 Example program without a TEMP specifier

COMMON T REAL,DIMENSION(1000,50) :: A,B,C,D . . . p DO J=1,50 p DO I=1,1000 p T=A(I,J)+B(I,J) p C(I,J)=T+D(I,J) p END DO p END DO . . . CALL SUB END

- 333 -

If it is known that the value of "T", at the end of the DO loop will not be referenced by the subroutine SUB, use TEMP specifier withT as shown in "Figure 12.32 Usage of TEMP specifier". This will improve the performance because the instructions that maintain thevalue of T at the end of the loop are no longer needed.

Figure 12.32 Usage of TEMP specifier

COMMON T REAL,DIMENSION(1000,50) :: A,B,C,D . . . !OCL TEMP(T) p DO J=1,50 p DO I=1,1000 p T=A(I,J)+B(I,J) p C(I,J)=T+D(I,J) p END DO p END DO . . . CALL SUB END

Note:

If variables which are not used for temporary purposes are specified in the TEMP specifier, the compiler will perform loop slicingincorrectly.

TEMP_PRIVATE(var[,var]...)

In the target loop of automatic parallelization, the TEMP_PRIVATE specifier directs the system to treat a specified variable var as alocal variable in each thread.

When the optimization message like the following that means automatic parallelization is prevented is output, specify this specifier toassign the target data to local area of each thread. As a result, the data dependency is solved, and the automatic parallelization ispromoted.

jwd5208p-i "file", line n : DO loop is not parallelized: reference to 'var' may be different from serial execution.

The TEMP_PRIVATE indicator has the following format:

!OCL TEMP_PRIVATE(var[,var]...)

Wildcards can be specified in var. See Section "12.2.3.2.5 Wild Card Specification" for information on specifying wildcards.

The TEMP_PRIVATE indicator can be entered in the loop position. It is effective for the corresponding DO loop at the loop position.

The variable that can be specified for this specifier is as follows.

- The scalar variable of integer type, real type or logical type.

- The array variable that meets the following conditions:

- Integer type, real type or logical type. And,

- The size of array variable has been fixed when compilation.

When this specifier is applied, the following message is output.

jwd5013p-i "file", line n : !OCL TEMP_PRIVATE is applied to variable 'var'.

See Section "12.2.3.3.4 Notes on Using Optimization Control Line" for the notes when specifying this specifier.

- 334 -

In "Figure 12.33 Example program without TEMP_PRIVATE specifier", this loop is not parallelized, because the order of array Adefinition and the reference is uncertain in an outside loop.

Figure 12.33 Example program without TEMP_PRIVATE specifier

INTEGER A(100), B(100), C(100,100) INTEGER M,N DO J=1,100 ! Target loop for automatic parallelization. DO I=1,M ! Because the value of variable M is uncertain, A(I) = B(I) ! the range where array A is defined is uncertain. ENDDO DO I=1,N ! Because the value of variable N is uncertain, C(I,J) = A(I) ! the range where array A is referred is uncertain. ENDDO ENDDO PRINT*, C END

In "Figure 12.34 Usage of TEMP_PRIVATE specifier", data dependency is solved by the TEMP_PRIVATE specifier. As a result,loop slicing is performed to the outside loop.

Figure 12.34 Usage of TEMP_PRIVATE specifier

!OCL TEMP_PRIVATE(A)p DO J=1,100p DO I=1,Mp A(I) = B(I)p ENDDOp DO I=1,Np C(I,J) = A(I)p ENDDOp ENDDO

FIRST_PRIVATE(var[,var]...)

In the target loop of automatic parallelization, the FIRST_PRIVATE specifier directs the system to treat a specified variable var as alocal variable in each thread. And, the initial value of the variable is directed to be referred at the loop entry of each thread.



The FIRST_PRIVATE indicator has the following format:

!OCL FIRST_PRIVATE(var[,var]...)


The FIRST_PRIVATE indicator can be entered in the loop position. It is effective for the corresponding DO loop at the loop position.



- 335 -





jwd5014p-i "file", line n : !OCL FIRST_PRIVATE is applied to variable 'var'.


In "Figure 12.35 Example program without FIRST_PRIVATE specifier", this loop is not parallelized, because the order of array Adefinition and the reference is uncertain in an outside loop.

Figure 12.35 Example program without FIRST_PRIVATE specifier

INTEGER A(100), B(100,100), C(100,100), D(100) INTEGER M,N DO J=1,100 ! Target loop for automatic parallelization. DO I=1,M ! Because the value of variable M is uncertain, B(I,J) = A(I) ! the range where array A is referred is uncertain. ENDDO DO I=N,100 ! Because the value of variable N is uncertain, A(I) = B(I,J) + D(I) ! the range where array A is defined is uncertain. C(I,J) = A(I) ENDDO ENDDO PRINT*, C END

In "Figure 12.36 Usage of FIRST_PRIVATE specifier", data dependency is solved by the FIRST_PRIVATE specifier. As a result,loop slicing is performed to the outside loop.

Figure 12.36 Usage of FIRST_PRIVATE specifier

!OCL FIRST_PRIVATE(A)p DO J=1,100p DO I=1,Mp B(I,J) = A(I)p ENDDOp DO I=N,100p A(I) = B(I,J) + D(I)p C(I,J) = A(I)p ENDDOp ENDDO

LAST_PRIVATE(var[,var]...)

In the target loop of automatic parallelization, the LAST_PRIVATE specifier directs the system to treat a specified variable var as alocal variable in each thread. And, the variable is directed to be referred after this DO loop.



- 336 -

The LAST_PRIVATE indicator has the following format:

!OCL LAST_PRIVATE(var[,var]...)


The LAST_PRIVATE indicator can be entered in the loop position. It is effective for the corresponding DO loop at the loop position.







jwd5015p-i "file", line n : !OCL LAST_PRIVATE is applied to variable 'var'.


In "Figure 12.37 Example program without LAST_PRIVATE specifier", this loop is not parallelized, because the order of array Adefinition and the reference is uncertain in an outside loop.

Figure 12.37 Example program without LAST_PRIVATE specifier

INTEGER A(100), B(100,100), C(100,100), D(100) INTEGER M,N DO J=1,100 ! Target loop for automatic parallelization. DO I=1,M B(I,J) = I ENDDO DO I=N,100 ! Because the value of variable N is uncertain, A(I) = B(I,J) + D(I) ! the range where array A is defined is uncertain. C(I,J) = A(I) ENDDO ENDDO PRINT *, A, C END

In "Figure 12.38 Usage of LAST_PRIVATE specifier", data dependency is solved by the LAST_PRIVATE specifier. As a result, loopslicing is performed to the outside loop.

Figure 12.38 Usage of LAST_PRIVATE specifier

!OCL LAST_PRIVATE(A)p DO J=1,100p DO I=1,Mp B(I,J) = Ip ENDDOp DO I=N,100p A(I) = B(I,J) + D(I)p C(I,J) = A(I)p ENDDOp ENDDO PRINT*, A, C

- 337 -

var1 op var2

var1 op cnst

These specifiers are used to indicate to the compiler the relationship between two variables or between a variable and a constant.

'var1' and 'var2' are variable names. 'cnst' is an integer constant or a named constant. Specify as follows:

!OCL var1 op var2!OCL var1 op cnst

The relational operators that can be specified for op are listed below.

Operator Explanation

.lt. and < Less than

.le. and <= Less than or equal to

.eq. and == Equal to

.ne. and != Not equal to

.gt. and > Greater than

.ge. and >= Greater than or equal to

The 'var1 op var2' or 'var1 op cnst' indicator can be entered in the loop position and total position.




In "Figure 12.39 Usage of the var1 op var2 or var1 op cnst specifier", parallelization is not performed when the OCL is not specifiedbecause it is not certain that the definition or order of data references of array A remains unchanged by parallelization, unless themagnitude relation between the array indexes is identified. op specifier enables parallelization by directing that the definition or orderof data references of array A remain unchanged before and after the parallelization.

Figure 12.39 Usage of the var1 op var2 or var1 op cnst specifier

REAL,DIMENSION(2000) :: A,B !OCL M.gt.2000 p DO I = 1,2000 p A(I) = A(I+M) + B(I) p END DO END

12.2.3.2.5 Wild Card Specification

In the operand of the following optimization control specifiers, a wild card may be specified for a variable name or a procedure name:

- INDEPENDENT

- TEMP

Wildcard is a combination of alphanumeric characters and a special character called a wildcard. Specifying a wildcard is equivalent tospecifying multiple variable names or procedure names that correspond to the criteria specified by the string specified as the wildcard.

The two wildcard characters "*" and "?" have the following meanings:

- "*" matches a string of one or more characters consisting of any alphanumeric characters.

- 338 -

- "?" matches any single alphanumeric character.

Note that multiple wildcards cannot be specified when specifying one wildcard.

!OCL TEMP(W*)

In this example, "W*" specifies a variable names where the first character is "W" and the length is two or more.

For example, the variable names WORK1, W2, and WORK3 match the wildcard.

!OCL INDEPENDENT(SUB?)

In this example, "SUB?" specifies a procedure name whose first three characters are "SUB" and the length is four.

For example, the array name SUB1, SUB2, and SUB9 match the wildcard.

12.2.3.3 Notes on Automatic ParallelizationThis section provides notes about using Automatic Parallelization.

12.2.3.3.1 Caution when specifying compile-time option -Kparallel,instance=N

When specifying the number of threads for parallel processing with the compiler option -Kparallel, instance=N, the N must be the sameas the number of threads used on runtime. See Section "12.2.1.2.2 Number of Threads" for information about specifying the number ofthreads.

When an incorrect value is specified on -Kinstance, the program will abort with the following runtime message.

jwe1040i-s This program cannot be executed, because the number of threads specified by -Kinstance=N option and the actual number of threads are not equal.

12.2.3.3.2 Nested Parallel Processing

Nested parallel processing is when a procedure is called from a sliced DO loop, and the procedure includes another sliced DO loop. Whenthis occurs, the inner DO loop will be executed serially. -Kparallel, instance=N must not be specified on compiling a source program thatmay cause nested parallel processing.

"Figure 12.40 Multiprocessing of nested loops" illustrates a sliced DO loop executed as a serial process. If a program with such a DO loopis compiled with the -Kparallel, instance=N option, the program may not work correctly.

Figure 12.40 Multiprocessing of nested loops

File: a.f

!OCL INDEPENDENT(SUB) DO I=1,100 ! Executed in parallel J=I CALL SUB(J) END DO : END RECURSIVE SUBROUTINE SUB(N) : DO I=1,10000 ! should be executed in serial A(I)=1/B(I)**N END DO : END

The program may not work correctly, if the source code 'a.f' is compiled with the following sample.

- 339 -

frtpx -Kparallel,instance=4 a.f (invalid use)

To prevent such errors, specify the optimization control line "!OCL SERIAL" for the procedure call defined in the DO loop for whichloop slicing is to be performed.

!OCL SERIAL RECURSIVE SUBROUTINE SUB(N) : DO I=1,10000 A(I)=1/B(I)**N END DO : END

12.2.3.3.3 Caution when specifying compile-time option -Kparallel,reduction

The execution result of a program compiled with -Kparallel,reduction may differ from one executed serially. This is because the sequenceof operation in parallel processing may be different from a serial one.

Example:

"Figure 12.41 Side effect of reduction optimization" is a sample used in Section "12.2.3.1.8 Loop Reduction". In serial processing, thearray element from A(1) to A(10000) will be sequentially appended to the variable SUM. In parallel processing, the sum of the arrayelement A(1) to A(5000) are stored as SUM1, A(5001) to A(1000) are stored as SUM2. SUM is then derived by adding SUM1 andSUM2.

In this way, the calculation sequence of the array A element in serial processing is different from parallel processing, which willintroduce a difference in precision.

Figure 12.41 Side effect of reduction optimization

SUM=0DO I=1,10000 SUM=SUM+A(I)END DO



SUM=SUM+SUM1+SUM2

12.2.3.3.4 Notes on Using Optimization Control Line

Invalid usage of the optimization control lines is shown blow.

The system will perform incorrect loop slicing because of the incorrect optimization control lines.

- TEMP specifier

The following program specifies TEMP incorrectly for variable T. There will not be a correct value assigned to variable LAST, becausethe system cannot guarantee a correct value of variable T at the end of the DO loop.

!OCL TEMP(T) DO I=1, 1000 T=A(I)+B(I) C(I)=T+D(I)

- 340 -

END DO LAST = T

- INDEPENDENT specifier

The following program specifies INDEPENDENT incorrectly for procedure SUB. The execution result is unpredictable when arrayA is sliced, because the order of the data references for array A in parallel execution is different from the order of data references inserial execution.

!OCL INDEPENDENT(SUB) DO I=2, 1000 A(I)=B(I)+1.0 CALL SUB(I-1) END DO ... END RECURSIVE SUBROUTINE SUB(J) COMMON A(1000) A(J)=A(J)+1.0 END

- TEMP_PRIVATE specifier

The following program specifies TEMP_PRIVATE incorrectly for variable A. Because an uncertain value is referred in each thread,the execution result is incorrect.

!OCL TEMP_PRIVATE(A) DO I=1,1000 B(I) = A ENDDO

- LAST_PRIVATE specifier

The following program specifies LAST_PRIVATE incorrectly for array A. Because the array A may not be defined by the last thread,the execution result is incorrect.

!OCL LAST_PRIVATE(A) DO J=1,1000 DO I=J,500 A(I) = B(J) ENDDO ENDDO PRINT *, A

12.2.3.3.5 I/O Statement and Intrinsic Procedure Call on Parallel Processing

When a sliced DO loop calls a procedure that includes an input-output statement, an intrinsic subroutine, or an intrinsic function that isnot supported by loop slicing, the program may not work correctly. The performance of the program may deteriorate because of theoverhead of parallel processing, or the result of the input/output statement may not be the same as with serial processing.

"Figure 12.42 Multiprocessing I/O statement" shows an example of input-output statements in a procedure called from a sliced DO loop.

Figure 12.42 Multiprocessing I/O statement

File: a.f

!OCL INDEPENDENT(SUB) DO I=1,100 J=I CALL SUB(J) END DO : END RECURSIVE SUBROUTINE SUB(N)

- 341 -

: PRINT *,N : END

12.2.4 Linkage with Other Multi-Thread ProgramsThis section describes the restrictions of linking with object program that perform in multiple threads rather than automatic parallelization.

OpenMP program by this Compiler

OpenMP object modules, which are compiled with the -Kopenmp option, can be linked with object programs that are compiled withthe -Kparallel option.

Other Multi-Thread Programs

Linking with object programs that use multiple threads, but which are not compiled with the system, is not supported.

However, the program may be linked with a pthread program when using the environment variable FLIB_PTHREAD.

Environment Variable FLIB_PTHREAD

You can control the linking with a pthread program, using the environment variable FLIB_PTHREAD.

Valid values are as follows. Default value is 0.

Value Explanation

0(Default)

The program can be linked with a pthread program which uses the pthread thread as a controlthread.

However, there are the following restrictions.

- The pthread thread must use suspending wait.

- The pthread program must not be compiled with -Kopenmp option or -Kparallel option.(*)

- The routine which is compiled with -Kopenmp option or -Kparallel option must not beused in the pthread program. (*)

- The Fortran routine must not be used in the pthread program. (*)

1 The program can be linked with a pthread program which executes parallel processing.

And, the program can be linked with a pthread program which uses the pthread thread as acontrol thread.


- A pthread parallel processing, and an OpenMP or automatic parallel processing must beexecuted in order. (*)An OpenMP or automatic parallel processing must not be executed in a pthread parallelprocessing.And, a pthread parallel processing must not be executed in OpenMP or automatic parallelprocessing.





- 342 -

Value Explanation

- Fujitsu's math libraries must not be used in the pthread program. (The pthread programmust not be compiled with -SSL2BLAMP option.) (*)

- You must execute OpenMP parallel processing with two or more threads, before creatingpthread threads. And, this number of OpenMP threads cannot be changed later.

And, when this function is used, the following functions are effective.

- FLIB_SPINWAIT=0

- FLIB_CPUBIND=off

Operation is not guaranteed when a value which is different in the above-mentionedenvironment variable is specified.

*) Operation is not guaranteed when the restriction function is used.

12.3 Parallelization by OpenMP SpecificationsThis section explains the parallelization based on the OpenMP specifications provided by this compiler.

This compiler is designed to compile programs that conform to the OpenMP Application Program Interface Version 3.1 July 2011.

See the "OpenMP Architecture Review Board" website for details about OpenMP Application Program Interface Version 3.1 July 2011.

12.3.1 Compilation and ExecutionThis section describes how to compile and execute the source program created on OpenMP specifications.

12.3.1.1 CompilationSpecify the compiler option -Kopenmp to compile and link a source program created to OpenMP specifications.

12.3.1.1.1 Compiler Options to Compile OpenMP Programs

This section explains the compiler options for OpenMP Programs.

-Kauto

Local variables, other than those with the SAVE attribute or initial values, are treated as automatic variables and allocated to the stackarea. However, the local variables of the main program are not allocated to the stack. See Section "9.11 Effects of Allocating onStack" for points to note about the size of the stack area.

This option is automatically set when the -Kopenmp option is specified.

-Knoauto

Local variables, other than those with the SAVE attribute or initial values, are not treated as automatic variables.

If this option is to be specified at the same time as the -Kopenmp option, it must be specified after the -Kopenmp option. Specifyingthis option with -Knothreadsafe option reduces the stack area size of the procedure being compiled.

This option can be specified for the main program or a subprogram called outside of the dynamic parallel region, and must be withOpenMP directives other than the following:

- PARALLEL constructs that do not include private variable declarations

- THREADPRIVATE directing statement

If this is specified for a procedure for which it cannot be specified, the program may not work correctly.

This option is effective only when it is specified with the -Knothreadsafe option. This is because, as with -Kauto, the -Kthreadsafeoption set by -Kopenmp allocates local variables to the stack.

- 343 -

-Kopenmp

-Kopenmp will activate the directives that conform to the OpenMP specifications. The -Kthreadsafe, -Kauto, and -D_OPENMP=201107 options are also activated when -Kopenmp is in effect. However, the local variable of the main program willnot be allocated to the stack even if the -Kauto option is set. The -Knoauto and -Knothreadsafe options can be valid only when theyare specified after the -Kopenmp option. Caution should be taken when the -Knoauto and -Knothreadsafe options are specified at thesame time as the -Kopenmp option (Refer to -Knoauto and -Knothreadsafe options). Even when only object programs are specifiedas file names (that is, when only linking is to be performed), this option must be specified if an object program that was compiled usingthe -Kopenmp option is included.

$ frtpx a.f90 -Kopenmp -c$ frtpx a.o -Kopenmp

-Knoopenmp

When the -Knoopenmp option is specified, the source program is compiled with the directives of OpenMP specifications treated ascomment lines.

When both -Kopenmp and -Knoopenmp are not specified, the compiler assumes that -Knoopenmp is specified.

-Kopenmp_ordered_reduction

The -Kopenmp_ordered_reduction option specifies to fix the operation order of the reduction operations same as in the numerical orderof the threads at the end of the region for which the reduction clause was specified. Note that execution performance may decreasecompared to when the -Kopenmp_ordered_reduction is invalidated. This option is valid only when the -Kopenmp option is in effect.

-Kopenmp_noordered_reduction

The -Kopenmp_noordered_reduction option specifies not to fix the operation order of the reduction operation at the end of the regionfor which the reduction clause was specified.

-Kopenmp_tls

The -Kopenmp_tls option directs that threadprivate variables are allocated to the Thread-Local Storage. This option is effective onlywhen the -Kopenmp option is set.

If the threadprivate variable is passed between Fortran and C/C++ programs, it is necessary to specify the -Kopenmp_tls option.

If the -Kopenmp_tls option is used, all the source programs comprising the executable program must be compiled with the -Kopenmp_tls option.

-Kopenmp_notls

The -Kopenmp_notls option directs that threadprivate variables are allocated to the heap.

-Kthreadsafe

A thread-safe object program is generated when the -Kthreadsafe option is specified. See Section "9.11 Effects of Allocating onStack" for points to note about the size of the stack area.

This option is automatically set when the -Kopenmp option is specified.

-Knothreadsafe

Object programs that are not thread-safe are created when the -Knothreadsafe option is specified.

If this option is to be specified at the same time as the -Kopenmp option, is must be specified after the -Kopenmp option. When thisoption is specified the instructions generated are not those that guarantee the process to be thread-safe.

This option can be specified for the main program or a subprogram called outside of the dynamic parallel region, and must be withOpenMP directives other than the following:

- PARALLEL constructs that do not include private variable declarations

- THREADPRIVATE directives

If this is specified for a procedure for which it cannot be specified, the program may not work correctly.

- 344 -

-Nprivatealloc

For a list item with the ALLOCATABLE attribute in a private clause of OpenMP: if the list item is "currently allocated", the new listitem will have an initial state of "not currently allocated" the program does not conform to the OpenMP 3.1 standard. This option isvalid only when the -Kopenmp option is in effect.

-Nnoprivatealloc

When an allocatable array is described in the OpenMP private section and the status is "currently allocated", a new list item will havean initial state of "currently allocated".

12.3.1.1.2 Displaying Optimization Information of OpenMP Program

This section describes the display of optimization information of an OpenMP program, when the compiler option -Qp or -Nlst=p isspecified. See Section "4.1.2 Compilation Information" for information about the compiler option -Qp and -Nlst=p.

When the compiler option -Qp or -Nlst=p is specified, the optimization information for the instructions generated by the OpenMP directivesare marked as follows:

- "p" is marked for statements that can be executed in parallel processing. It is not marked for the statements executed redundantly.Statements directly enclosed in a DO or SECTIONS statement will meet this condition. Within the WORKSHARE directive, onlythe parallelized statements are marked as "p".

- "s" is marked for statements that will be executed in only one thread. Statements directly enclosed in MASTER, SINGLE, CRITICAL,or ORDERED directive will meet this condition. For the instructions just after the ATOMIC directive, "s" will be marked only whenwhole instructions are executed within a single thread.

- "m" is marked for statements which include both instructions that will be executed in parallel and in a single thread. A statementimmediately after an ATOMIC directive within a WORKSHARE directive may meet this condition, for example.

For nested parallelization, only the parallelization information of the innermost parallel region, which includes the target instructions, isdisplayed. Whether the parallel region is executed in parallel or serially does not affect the display.

12.3.1.2 Execution ProcessThe procedure to run the OpenMP program is the same as executing it as a serial process.

12.3.1.2.1 Environment Variable at ExecutionTHREAD_STACK_SIZE

A user can specify the stack area size for each thread using the environment variable THREAD_STACK_SIZE in Kbytes. When theenvironment variable OMP_STACKSIZE is specified, the greater value is used as the stack area size for each thread. See Section"12.3.1.2.3 Notes during Execution" for details.

OMP_STACKSIZE

A user can specify the stack area size for each thread using the environment variable OMP_STACKSIZE in Kbytes, Mbytes, or Gbytes.When the environment variable THREAD_STACK_SIZE is specified, the greater value is used as the stack area size. See Section"12.3.1.2.3 Notes during Execution" for details.

OMP_PROC_BIND

A user can control the CPU Binding for Thread using the environment variable OMP_PROC_BIND.

Valid values and their meanings are as follows:

TRUE : The thread is bound to a CPU.

FALSE : The thread is not bound to a CPU.

The environment variable FLIB_CPUBIND or FLIB_CPU_AFFINITY is given to priority more than OMP_PROC_BIND.

However, OMP_PROC_BIND is effective in the following case.

- FLIB_CPUBIND=off is specified, and FLIB_CPU_AFFINITY is not specified.

- 345 -

When TRUE is specified, threads are bound to CPUs in order of cpuid list. However, cpuid which can actually be used becomes onlywithin the limits of CPU affinity of the process at the time of an execution start.

See Section "D.2.1 FLIB_CPUBIND" for details.

See Section "12.3.1.2.3 Notes during Execution" for details.

FLIB_SPINWAIT and OMP_WAIT_POLICY

The standby threads during the serial execution can be controlled with the environment variable FLIB_SPINWAIT orOMP_WAIT_POLICY.

FLIB_SPINWAIT value:

unlimited : Uses spin wait.

0 : Uses suspending wait.

<n>s : Uses spin wait until after <n> seconds, and uses suspending wait.

<n>[ms] Uses spin wait until after <n> milliseconds, and uses suspending wait.

<n> is integer value that must be positive.

The default is unlimited.

Spin waiting is a technique to wait while consuming CPU. Suspend waiting is a technique to wait without consuming CPU. Togive priority to elapsed time, specify unlimited, as spin waiting requires less overhead for parallel process. Select 0 to give priorityto the CPU time.

OMP_WAIT_POLICY value:

ACTIVE : Uses spin wait.

PASSIVE : Uses suspending wait.

The default is "ACTIVE".

Select ACTIVE to give priority to the elapsed time. Select PASSIVE to give priority to the CPU time.

The settings of environment variable FLIB_SPINWAIT will be valid when environment variable FLIB_SPINWAIT is specified.

FLIB_FASTOMP

A user can control the use of the high speed runtime library using the environment variable FLIB_FASTOMP. Valid values and theirmeanings are as follows: The default is "TRUE".

TRUE : High speed runtime library is used.

FALSE : High speed runtime library is not used.

The high-speed runtime library achieves its speed by restricting part of the OpenMP specification. The features of the high-speedruntime library are as follows.

- This assumes that the program does not have nested parallelization, and one CPU is dedicated to one thread.

- The hardware barrier can be used in Job Operation Software environment.

The following restrictions are applied to use the high-speed runtime library:

- The number of threads of the OMP_SET_NUM_THREADS subroutine, or the number of threads specified in theNUM_THREADS clause in the PARALLEL directive, must be the same as the number of threads determined in the runtimeenvironment.

See Section "12.3.1.2.3 Notes during Execution" for the number of threads determined in the runtime environment.

The following restrictions are applied for controlling the execution of the OpenMP environment variable and the service procedurewhen using the high-speed runtime library.

- Regardless of the configuration defined in the environment variable OMP_NESTED and the configuration of OMP_SET_NESTEDsubroutine, nested parallel regions are always serialized.

- 346 -

- The number of threads cannot exceed the number of available CPUs, regardless of the configuration defined in the environmentvariable OMP_DYNAMIC and the configuration of OMP_SET_DYNAMIC subroutine.

See Section "12.3.1.2.3 Notes during Execution" for information on the number of available CPUs.

12.3.1.2.2 Environment Variable for OpenMP Specifications

The following environment variables are supported by the OpenMP specifications. See the "OpenMP specifications" for details.

OMP_SCHEDULE

This specifies the schedule type and the chunk size that should be performed for the DO and PARALLEL DO directives with scheduletype RUNTIME.

OMP_NUM_THREADS

This specifies the number of threads used during execution.

OMP_DYNAMIC

This specifies to enable and disable dynamic thread adjustment.

OMP_PROC_BIND

This specifies whether threads are bound to CPUs.

OMP_NESTED

This enables or disables nested parallelism.

OMP_STACKSIZE

This specifies the stack size for each thread.

OMP_WAIT_POLICY

This specifies the behavior of the waiting thread.

OMP_MAX_ACTIVE_LEVELS

This specifies the maximum number of nested active parallel regions.

OMP_THREAD_LIMIT

This specifies the maximum number of threads executed on an Open MP program.

12.3.1.2.3 Notes during Execution

The following points should be noted when using parallelization features based on the OpenMP specifications on this compiler.

Allocating Variables at Runtime

The programs on this system that use parallelization features based on OpenMP specifications allocate local variables (other than theautomatic data object while compiler option -Knoautoobjstack is in effect) and private variables on the stack, except for the main program.When a procedure requires a large area for local and private variables, a sufficiently large stack must be allocated.

The maximum process stack area can be specified using a command such as "ulimit" (bash built-in command).

The stack area size for each thread is the same as that for the process. When the maximum stack size for the process is larger than min( thesize of available memory in this system, the size of a virtual memory ) / ( the number of threads )

the system will determine the stack area size as follows and allocate it. However, the assumed value does not guarantee that the systemworks correctly. It is recommended that an appropriate value for maximum process stack size be specified manually. Note that the maximumvirtual memory size can also be confirmed using ulimit (bash built-in command).

S = (M/T)/5

S: the stack size for each thread (byte)M: a size with a smaller mounting memory and a smaller virtual memory (byte)T: the number of threads created in the first parallel region

- 347 -

The compiler calculates the stack area size without considering the feature that can dynamically change the thread numbers to align thestack area size of each thread. For this reason, the formula presented above uses the thread number assuming "FLIB_FASTOMP=TRUE".When the thread number is to be changed dynamically, specify the maximum thread numbers using OMP_NUM_THREADS in advance.

When the stack area size for each thread needs to be specified, use environment variable THREAD_STACK_SIZE or OMP_STACKSIZE.

See Section "12.3.1.2.1 Environment Variable at Execution" for information about environment variables THREAD_STACK_SIZE andOMP_STACKSIZE.

Maximum number of CPUs

The maximum number of CPUs is as follows:

- For jobs:Number of CPUs that can be used with jobsSee Section "D.1 Management of the CPUs resource" for details.

- Not for job:Number of CPUs in the system

Number of Threads

The number of threads is common for both OpenMP and automatic parallelization when using the high-speed runtime library, otherwisethe number will be determined independently. Use of the high-speed runtime library is determined by the environment variableFLIB_FASTOMP (See Section "12.3.1.2.1 Environment Variable at Execution").

- When using the high-speed runtime library (i.e., when FLIB_FASTOMP is TRUE), the thread number will be determined with the following priority:

1. The value specified in the environment variable OMP_NUM_THREADS (Note)

2. The value specified in the environment variable PARALLEL


4. 1 thread

(Note)

When -Kparallel is specified during the linkage phase of compilation, the value must be the same as the environment variablePARALLEL if specified. When the value does not match, the following message will be output, and the smaller value will beused.

jwe1042i-w The value specified in the environment variable PARALLEL and OMP_NUM_THREADS are different when the environment variable FLIB_FASTOMP is TRUE. The value num is used as the value specified in the environment variable.

If the number of threads determined in this priority exceeds the maximum number of CPUs, the number of threads will be alignedto the maximum number of CPUs.

The value specified in the NUM_THREADS clause section of the PARALLEL directive or defined in theOMP_SET_NUM_THREADS service routine must be the same as the number of threads determined here. If a different value isspecified, the program will terminate with the following message.

jwe1041i-s The number of thread cannot be changed when the environment variable FLIB_FASTOMP is TRUE.

- When the high-speed runtime library is not used (i.e. when FLIB_FASTOMP is FALSE), the number of threads for OpenMP will be determined using the following priority. The number of threads for automatic parallelizationfollows the description in Section "12.2 Automatic Parallelization".

1. The value specified in the NUM_THREADS clause in the PARALLEL statement

2. The value specified in the OMP_SET_NUM_THREADS service routine.

3. The value specified in the environment variable OMP_NUM_THREADS

- 348 -

4. The value specified in the environment variable PARALLEL


6. 1 thread

When the dynamic thread adjustment feature is enable, the smaller value of the number of threads determined by the above priority, andthe CPU number will be used as the number of threads.

When the dynamic thread adjustment feature is disable, the thread number determined by the above priority will be used, even if it exceedsthe maximum number of CPUs. When more than one thread is assigned for a CPU, the thread that is to be executed in parallel will beexecuted in a time sharing manner. In this case, the overhead caused by synchronizing the threads may sacrifice the performance of theprogram. It is recommended to conserve resources by setting one thread per CPU, unless there is a specific reason to do otherwise. SeeSection "D.1 Management of the CPUs resource" for information on the number of CPUs that can be used with jobs.

Notes when setting time limits

The program may not work correctly when time limits are set using the command such as ulimit(1). In such a case, use the Fortran exceptionhandling process, execution time option -Wl, -i to invalidate it.

Alternatively, use the execution time option -Wl,-t to set the time limit.

Notes when using service routines

- ALARM Service Function

It may not work correctly. It is only for a program using serial processing.

- KILL and SIGNAL Services Functions

It may cause a deadlock. It is only for a program using serial processing.

- FORK Service Function

It may not work correctly. It is only for a program using serial processing.

- SYSTEM, SH and CHMOD Service Functions

Memory usage doubles and performance may deteriorate. Use only with programs using serial processing.

CPU Binding for Thread

CPU Binding for Thread differs between job execution and non-job execution.

When executing program on non-job.

The thread is not bound to a CPU. But you can control CPU Binding for Thread using environment variableFLIB_CPU_AFFINITY.

The environment variable FLIB_CPU_AFFINITY

Threads are bound to CPUs in order of the specified cupid list.

When the number of specified CPUs is exceeded, it is repeatedly used from the beginning of the list.

The cpuid shall be separated by comma (',') or space(' ').

The cpuid list can have the next form that has range with increment.

cpuid1[-cpuid2[:inc]]

cpuid1: cpuid of the beginning of the range. (0<=cpuid1<CPU_SETSIZE)

cpuid2: cpuid of the end of the range. (0<=cpuid2<CPU_SETSIZE)

inc: increment (1<=inc<CPU_SETSIZE)

In addition, it is necessary to be the following.

cpuid1<=cpuid2

- 349 -

It becomes equivalent to the case where all CPUs for every increment value inc in the range from cpuid1 to cpuid2 are

specified.

The cpuid can be used the above-mentioned value. However, cpuid which can actually be used becomes only within the limits ofCPU affinity of the process at the time of an execution start.

See CPU_SET(3) about details of CPU_SETSIZE.

Examples of using environment variable FLIB_CPU_AFFINITY are shown below.

Example 1:

$ export FLIB_CPU_AFFINITY="0,2,1,3"

The thread is bound to CPU in order of 0, 2, 1, and 3.When the number of threads is five or more, it is repeatedly used from the beginning of the list.

Example 2:

$ export FLIB_CPU_AFFINITY="0-7"

The thread is bound to CPU in order of 0, 1, 2, 3, 4, 5, 6 and 7.When the number of threads is nine or more, it is repeatedly used from the beginning of the list.

Example 3:

$ export FLIB_CPU_AFFINITY="0-7:2"

The thread is bound to CPU in order of 0, 2, 4 and 6.When the number of threads is five or more, it is repeatedly used from the beginning of the list.

Example 4:

$ export FLIB_CPU_AFFINITY="0-4:2,1,7"

The thread is bound to CPU in order of 0, 2, 4, 1 and 7.When the number of threads is six or more, it is repeatedly used f from the beginning of the list.

When executing program on job.

When thread is not bound to CPU, you may control CPU Binding with FLIB_CPU_AFFINITY as executing on non-job.

When thread is bound to CPU on job, the environment variable FLIB_CPU_AFFINITY becomes invalid.

For details of CPU Binding, see Section "D.2 CPU Binding".

12.3.1.2.4 Output from Multiple Threads

If errors are detected from multiple threads in a program while it is running, diagnostic message and trace back maps are output for eachthread. For this reason, a piece of information per thread is output in a parallel region.

12.3.2 Implementation-Dependent SpecificationsThe implementation-dependent specification in the OpenMP is implemented in the following way in this compiler. Refer to the OpenMPspecifications for details about these items.

Memory Model

When multiple threads attempt to update a variable that exceeds 4 bytes or is stored across a 4-byte boundary, the value of the variablemay be left undefined without storing any of the values, unless lock control is explicitly performed.

If read and write to the variable occurs from multiple threads at the same time, and this exceeds 4 bytes or crosses a 4-byte boundary,the read value may be undefined rather than the written value, unless lock control is explicitly performed.

Internal Control Variables

The initial values for each of the internal control variables are as follows:

- nthreads-var: 1.

- 350 -

- dyn-var: TRUE.

- run-sched-var: STATIC without chunk size.

- def-sched-var: STATIC without chunk size.

- bind-var:FALSE.

- stacksize-var: the stack area size for threads in Section "12.3.1.2.3 Notes during Execution".

- wait-policy-var: ACTIVE.

- thread-limit-var: 2147483647.

- max-active-levels-var: 2147483647.

Dynamic Thread Adjustment Features

This compiler provides dynamic thread adjustment for parallelization performed based on OpenMP specifications. See Section"12.3.1.2.3 Notes during Execution" for information about the effect of this feature.

The dynamic thread adjustment feature is set by default.

Loop Statement

An eight byte integer type is used to calculate the iteration count of a collapsed loop.

When AUTO is set for the internal control variable run-sched-var, the effect of the SCHEDULE(RUNTIME) construct is set toSCHEDULE(STATIC).

SECTIONS Construct

Assignment to a thread of a structured block in a SECTIONS construct is performed in the same way as a dynamic schedule. When athread reaches the SECTIONS region or finishes performing one of the structured blocks in the region, the thread starts performingthe next structured block in this sequence.

SINGLE Construct

SINGLE region is executed by the thread that reaches the region first.

Task scheduling points

In untied TASK regions, task scheduling points are located at the same positions as in the case of tied TASK regions: only in TASK,TASKWAIT, explicit or implicit BARRIER constructs, and at the completion point of the task.

ATOMIC Construct

Two ATOMIC regions will be executed independently (not exclusively), when a variable type to be updated or a type parameter isdifferent. When the type and type parameter match, it may be executed exclusively even if the address is different.

- When updating a logical type, complex number type, 1 byte integer type, 2 byte integer type, or quadruple precision (16 byte) realtype variable

- When updating array element and index expression is not the same

- The target expression has explicit or implicit type conversion

OMP_SET_NUM_THREADS Routine

Calling the OMP_SET_NUM_THREADS routine has no effect when the argument value is equal to or less than 0. A value that exceedsthe number of threads supported by the system must not be specified.

OMP_SET_SCHEDULE Routine

There are no schedule types that are implementation-dependent.

OMP_SET_MAX_ACTIVE_LEVELS Routine

A call to the OMP_SET_MAX_ACTIVE_LEVELS routine will be ignored when it is performed from an explicit PARALLEL region.It will also be ignored when the argument is an integer that less than 0.

OMP_GET_MAX_ACTIVE_LEVELS Routine

The OMP_GET_MAX_ACTIVE_LEVELS routine can be called from anywhere in the program and it returns the value of the internalcontrol variable max-active-levels-var.

- 351 -

Environment Variable OMP_SCHEDULE

When the schedule type specified for OMP_SCHEDULE is invalid, the schedule type is ignored, and the default schedule (STATICwithout chunk size) is used.

When the schedule type specified for OMP_SCHEDULE is STATIC, DYNAMIC, or GUIDED, and the chunk size is not a positivenumber, the chunk size will be as follows:

STATIC : no chunk size

DYNAMIC, orGUIDED

: 1

Environment Variable OMP_NUM_THREADS

When a value equal to or less than 0 is specified for the list of OMP_NUM_THREADS, it works in the same ways as when 1 isspecified. A value that exceeds the number of threads supported by the system must not be specified.

Environment Variable OMP_PROC_BIND

When a value other than TRUE or FALSE is specified for OMP_PROC_BIND, the value is ignored, and the default value (FALSE)is used.

Environment Variable OMP_DYNAMIC

When a value other than TRUE or FALSE is specified for OMP_DYNAMIC, the value is ignored and the default value (TRUE) isused.

Environment Variable OMP_NESTED

When a value other than TRUE or FALSE is specified for OMP_NESTED, it is ignored and the default value (FALSE) is used.

Environment Variable OMP_STACKSIZE

When the value specified for OMP_STACKSIZE does not meet the defined format, it is ignored and the default value (Refer to"12.3.1.2.3 Notes during Execution") is used.

Environment Variable OMP_WAIT_POLICY

ACTIVE performs spin wait. PASSIVE performs suspend wait.

Environment Variable OMP_MAX_ACTIVE_LEVELS

When the value specified for OMP_MAX_ACTIVE_LEVELS is an integer that is less than 0, it is ignored and the default value(2147483647) is used.

Environment Variable OMP_THREAD_LIMIT

When the value specified for OMP_THREAD_LIMIT is not a positive integer, it is ignored and the default value (2147483647) isused.

THREADPRIVATE directing statement

The value, allocation state, and association state of the thread private variables will be maintained across the consecutive PARALLELregions, only when the following conditions are satisfied. These conditions are not affected by the configuration of the dynamic threadadjustment feature and nested parallelization.

- The consecutive PARALLEL region is not nested within another explicit PARALLEL region.

- The same number of threads is used to execute both PARALLEL regions.

If these conditions are not satisfied, the value, allocation status, and association status of the threadprivate variable at the entry pointof the subsequent PARALLEL region will have an undefined status. However, the allocation status will be not allocated at the entrypoint of the subsequent PARALLEL region, when the following is true:

- The allocation status is not allocated just before entering the subsequent PARALLEL region.

- The allocation status is not allocated by any thread at the exit of the all preceding PARALLEL regions that are already executed.

- 352 -

SHARED Clause

When passing a shared variable to a non-intrinsic procedure, the values of the shared variables are stored in a temporary storage area,and the values are stored in the actual argument area after the procedure call. This can occur when the following conditions are satisfied:

a. The actual argument is one of the following:

- A shared variable

- A subobject of a shared variable

- An object associated to a shared variable

- An object associated to a subobject of shared variable

b. In addition the actual argument is one of the following:

- An array section

- An array section with a vector subscript

- An assumed-shape array

- A pointer array

c. The associated dummy argument for this actual argument is an explicit-shape array or an assumed-size array.

Runtime Library Definitions

This compiler provides an include file omp_lib.h and a module omp_lib for the parallelization feature based on OpenMP specifications.This does not include an interface block with generic name.

The following notes must be considered when using the module omp_lib and omp_lib.h:

- When the compiler option -AU is specified, module name omp_lib and each library routine name must be specified using lowercase alphabetic characters.

- When the compiler option -CcdII8, -CcI4I8, -CcdLL8, -CcL4L8, -Ccd4d8, or -Cca4a8 is specified, the corresponding type of thelibrary routine argument and the result will be converted. On execution, the corresponding runtime option -Lb or -Li must bespecified.

- Although the diagnostic message jwd2004i-i will be displayed if the named constant declared in omp_lib.h was not used, it doesnot affect the result.

12.3.3 Clarifying OPENMP Specifications and restrictionsThis section explains the interpretation of the OpenMP Specifications and some restrictions in this system.

12.3.3.1 Specifying Label with ASSIGN Statement and Assigned GO TO StatementAn ASSIGN statement cannot reference a label outside of the structured block that the ASSIGN statement itself belongs to. An ASSIGNstatement cannot reference a label in a PARALLEL region from outside of the PARALLEL region.

An assigned GO TO statement must not jump into, or jump out of the structured block range.

12.3.3.2 Additional Functions and Operators in ATOMIC Directive and REDUCTIONClause

The following intrinsic functions and operations can be specified in an ATOMIC directive and REDUCTION clause.

Intrinsic function : AND, OR

Operator : .XOR., .EOR.

12.3.3.3 Notes on Using THREADPRIVATEWhen using the THREADPRIVATE directive, the same common block name must be used for all program units and procedures in theprogram specified on THREADPRIVATE.

- 353 -

The common block, specified with THREADPRIVATE, cannot have the size extended.

12.3.3.4 Inline ExpansionThe following procedures are not expanded inline.

- User defined procedures which include OpenMP directives.

12.3.3.5 Internal Procedure Calling from Parallel RegionAll the parent procedure variables, which are referenced in the internal procedure called from a parallel region, are regarded as SHARED,even if they are privatized in the parallel region.

Example:

... I=1 THIS "I" IS SHARED.!$OMP PARALLEL PRIVATE(I) I=2 * PRINT *,I *"I" IS PRIVARE IN THIS RANGE. CALL C_PROC() *!$OMP END PARALLEL CONTAINS SUBROUTINE C_PROC() ... * PRINT *,I *"I" is SHARED IN THIS RANGE ... * END SUBROUTINE ...

12.3.3.6 Polymorphic VariablePolymorphic variable cannot be specified in THREADPRIVATE directive, and unlimited polymorphic variable cannot be specified asfollowing clauses:

- PRIVATE

- FIRSTPRIVATE

- LASTPRIVATE

- COPYPRIVATE

- COPYIN

12.3.3.7 BIND attributeIf the -Kopenmp_tls option is invalid, a variable or common block declared with BIND attribute cannot be specified in THREADPRIVATEdirective.

12.3.3.8 OPTIONAL attributeWhen a variable declared with OPTIONAL attribute is specified in PRIVATE, FIRSTPRIVATE, LASTPRIVATE, COPYPRIVATE orCOPYIN clause, the variable cannot be specified as actual argument of PRESENT intrinsic function in the scope.

12.3.3.9 ASSOCIATE NAMEAn associate name of SELECT TYPE or ASSOCIATE construct cannot be specified in OpenMP directive.

12.3.3.10 SelectorThe variable of selector in an ASSOCIATE construct or the variable of selector with the associate name in a SELECT TYPE constructcannot be specified as following clauses:

- 354 -

- PRIVATE

- FIRSTPRIVATE

- LASTPRIVATE

- COPYPRIVATE

- COPYIN

- REDUCTION

12.3.3.11 Derived type variable with length type parameterThe derived type variable with length type parameter cannot be specified as following clauses:

- PRIVATE

- FIRSTPRIVATE

- LASTPRIVATE

- COPYPRIVATE

- COPYIN

12.3.4 Notes on Creating Program

12.3.4.1 Implementing PARALLEL Region and Explicit TASK RegionThe structured block within a PARALLEL construct or a TASK construct is compiled as an internal procedure.

The internal procedures generated from a PARALLEL construct will be named "_OMP_IdNumber_".

The internal procedures generated from a TASK construct will be named "_TSK_IdNumber_".

12.3.4.2 Automatic Parallelization for OpenMP ProgramsThis compiler allows the specification of the compiler options -Kopenmp and -Kparallel at the same time. When they are specified together,the DO loop automatic parallelization performed is restricted as follows:

- Automatic parallelization is not performed for DO loops within the OpenMP constructs.

- Automatic parallelization is not performed for DO loops that statically include OpenMP directives.

- When there is a DO loop which is parallelized by the OpenMP DO directive, automatic parallelization is not performed for the followingDO loops:

- DO loops parallelized by the OpenMP DO directive

- DO loops within a DO loop parallelized by the OpenMP DO directive

12.3.5 Linkage with Other Multi-Thread ProgramsThis section describes the restrictions to link with multi-threaded programs other than OpenMP.

Programs Automatically Parallelized Using this Compiler

Object modules compiled with this compiler using the -Kparallel option can be linked with other modules compiled with this compilerusing the -Kparallel option.

Other multi-thread programs

The object module created by this system cannot be linked with multi-threaded object modules created by another multithread system.

However, the program may be linked with a pthread program when using the environment variable FLIB_PTHREAD.

Environment Variable FLIB_PTHREAD

You can control the linking with a pthread program, using the environment variable FLIB_PTHREAD.

- 355 -

Valid values are as follows. Default value is 0.

Value Explanation

0(Default)

The program can be linked with a pthread program which uses the pthread thread as a controlthread.






1 The program can be linked with a pthread program which executes parallel processing.

And, the program can be linked with a pthread program which uses the pthread thread as acontrol thread.


- A pthread parallel processing, and an OpenMP or automatic parallel processing must beexecuted in order. (*)An OpenMP or automatic parallel processing must not be executed in a pthread parallelprocessing.And, a pthread parallel processing must not be executed in OpenMP or automatic parallelprocessing.





- Fujitsu's math libraries must not be used in the pthread program. (The pthread programmust not be compiled with -SSL2BLAMP option.) (*)

- You must execute OpenMP parallel processing with two or more threads, before creatingpthread threads. And, this number of OpenMP threads cannot be changed later.

And, when this function is used, the following functions are effective.

- FLIB_SPINWAIT=0

- FLIB_CPUBIND=off

Operation is not guaranteed when a value which is different in the above-mentionedenvironment variable is specified.

*) Operation is not guaranteed when the restriction function is used.

12.3.6 Debug OpenMP ProgramsWhere an error occurs, a program terminates abnormally, or a program does not provide an expected result, the source code must beanalyzed and corrected.

This section describes the debug features to analyze the problems. Note that this section describes the debugging only for the OpenMPprograms. Also refer to "Chapter 8 Debugging".

- 356 -

12.3.6.1 Checking Features for DebuggingThis section describes the restriction of debugging features.

- The compiler option -H, which provides the checking feature for debugging, is not effective within the parallel region (includingdynamic parallelization).

12.4 I/O Buffer Parallel TransferThe I/O buffer parallel transfer is a feature that provides parallel threading to improve the performance of data transfer between the bufferarea used in the array area and the input-output statement in a Fortran program.

12.4.1 Compilation and ExecutionThis section describes the procedures and criteria to compile and execute a program using I/O buffer parallel transfer.

12.4.1.1 CompilationTo perform I/O buffer parallel transfer, specify the compiler option -Kparallel or -Kopenmp.

The compile time option -Kparallel is an option to perform automatic parallelization. See Section "12.2 Automatic Parallelization" fordetails.

The compiler option -Kopenmp is an option to perform parallelization based on the OpenMP specifications. See Section "12.3Parallelization by OpenMP Specifications" for details.

12.4.1.2 Execution ProcessTo use I/O buffer parallel transfer, the following configuration is required:

- - Specify MP for the environment variable FLIB_IOBUFCPY.

- - Specify the number of threads using the environment variable PARALLEL or OMP_NUM_THREADS.

See Section "12.2.1.2.2 Number of Threads" for information on the number of threads.

12.4.1.3 Conditions to Perform I/O Buffer Parallel TransferThe following conditions must be met to perform I/O buffer parallel transfer:

- Target I/O statements are unformatted I/O statements

- Target I/O statements are outside the parallel region or outside the automatically parallelized area

- When the environment variable FLIB_IOBUFCPY_SIZE is not specified, the amount of the data transferred and the I/O buffer valuemust be more than "NumberOfThread + 30" Kbyte.

- When size is specified on the environment variable FLIB_IOBUFCPY_SIZE, the amount of the data transferred and the I/O buffervalue must be more than "size" Kbyte.

For sequential access I/O statements, specify the I/O buffer size using the runtime option -g or the environment variable fuxxbf. For directaccess I/O statements, specify using the recl specifier of the open statement and the runtime option -d.

See "3.3 Runtime Options" for more information on runtime option -g.

See Section "3.8 Using Environment Variables for Execution"for information on the environment variables fuxxbf andFLIB_IOBUFCPY_SIZE.

12.4.1.4 Compilation execution example

The used approximate stack size for each thread is displayed per Byte.$ cat test.f character ch((30+4)*1024) ch='X'

- 357 -

open(10,form="unformatted") write(10) ch close(10) end$ frtpx -Kparallel test.f$ export FLIB_IOBUFCPY="MP"$ export PARALLEL=4$ ./a.out -Wl,-g34

This specifies MP for the environment variable FLIB_IOBUFCPY to enable I/O buffer parallel transfer. It also sets 4 for the environmentvariable PARALLEL to make the number of threads 4. The runtime option -g34 sets the I/O buffer size to 34 Kbytes (Number of thread4 + 30 Kbytes).

12.4.2 Notes on I/O Buffer Parallel TransferWhen the I/O buffer size or the transferred data size is small, the performance may deteriorate even if I/O buffer parallel transfer isperformed.

- 358 -

Appendix A Limits and RestrictionsThe Fortran compiler restricts the size and complexity of source programs. Some restrictions apply to one Fortran statement and othersapply to a combination of several statements.

A.1 Nesting Level of Functions, Array Sections, Array Elements,and Substring References

In a combination of function, array section, array element, and substring references, nesting must not exceed 255 levels. Depending onthe compiler, nesting may be subject to further restrictions.

Example: Nesting level of functions and array element references

INTEGER IA(10)EXTERNAL FUN,IFUN :X=FUN(IFUN(IA(1))) ! The nesting level is 3

A.2 DO, CASE, IF, SELECT TYPE, BLOCK, CRITICAL, andASSOCIATE Constructs

In a combination of DO, CASE, IF, SELECT TYPE, BLOCK, CRITICAL, and ASSOCIATE constructs, the total nesting level must notexceed 50.

Example: Nesting levels of DO and IF constructs

DO I=1,10 A(I)=A(I)+1 ! The nesting level is 1 IF (I.LE.5)THEN CALL S! The nesting level is 2 DO J=1,5 B(J)=A(J)! The nesting level is 3 END DO CALL T! The nesting level is 2 END IF END DO

A.3 Implied DO-LoopsThe nesting level of implied DO-loops in input/output or DATA statements must not exceed 25.

Example: Nesting level of implied DO-loops

PRINT *, ((A(I,J),I=1,2),J=1,3) ! The nesting level is 2

A.4 INCLUDE LineThe nesting level of INCLUDE lines must not exceed 16.

- 359 -

Example: Nesting level of INCLUDE lines

In this case, the nesting level of INCLUDE lines is 1.

A.5 Array DeclarationsFor array declarations, there are two restrictions.

A.5.1 Restriction for All Array DeclarationsThe compiler calculates T for each array declaration to reduce the number of calculations needed for array sections or array elementaddresses. The absolute value of T obtained by following formula must not exceed the limitation value, and the absolute value must notexceed the limitation value during calculation:

n : Array dimension numbers : Array element lengthl : Lower bound of each dimensiond : Size of each dimensionT : Value calculated for the array declaration

The limitation value is 9223372036854775807.

Example: Restrictions on array declarations

PARAMETER (IL=10000000,IU=10000999)INTEGER(8) A(IL:IU,IL:IU,IL:IU,IL:IU,IL:IU)

Because T (the value calculated for array declarations) for array A is 80080080080080000000, this array declaration must not be specified.

A.5.2 Restriction for Array Declaration of Zero-Sized ArrayIn an array declaration of zero-sized array that has bounds that are nonconstant specification expressions, the absolute value of T obtainedby following formula must not exceed the limitation value, and the absolute value must not exceed the limitation value during calculation:

n : Array dimension numbers : Array element length

- 360 -

l : Lower bound of each dimensionu : Upper bound of each dimensionT : Value calculated for the array declaration

The limitation value is 9223372036854775807.

A.6 Array SectionIn the following subscript triplet,

subscript1 : subscript2 : stride

overflow must not occur during calculation.

(subscript2 - subscript1 + stride) / stride

A.7 Distance between a Branch Instruction and the BranchDestination Instruction

The immediate field size of the branch destination address in a branch instruction is restricted. For this reason, when the branch instructionis separated from the branch destination instruction over a distance that exceeds a certain fixed value, the assembler may not be able toassemble the program. In this case, the assembler displays the following messages:

/var/tmp/asmAAAa006jR.s: Assembler messages: /var/tmp/asmAAAa006jR.s:10: Error: relocation overflow

The likelihood for this symptom becomes high when the program is assembled with the -H and -Kthreadsafe options specifiedsimultaneously for a loop with several thousand or more executable statements or when there are numerous array expressions in a hugeloop.

If this symptom occurs, correct the program (e.g., perform loop splitting) or -O0 option is in effect. It is necessary to note it for the executionperformance because -O0 option suppresses the optimization.

- 361 -

Appendix B Printing Control CommandThe printing control command, fot(1) and fotpx(1) convert files output by Fortran programs into standard I/O files for printing on a lineprinter. You can use fot(1) and fotpx(1) to map from Fortran output to the printer. That is, they map files or pipes that have Fortran carriage-control conventions into a form acceptable to conventional printers.

The first character of a formatted record provides the following control information.

Character Action

Blank Advance to the next line and print the line.

0 Advance two lines and print the line.

1 Advance to the first line of the next page and print the line.

+ Do not advance to the next line before printing the line (over-print the current line).

If a formatted record is empty, the first character of the line is regarded as the control character blank.

If the first character of a formatted record is neither a blank, 0, 1, nor +, a diagnostic message is output, and the first character is regardedas the control character blank.

B.1 Command FormatThe format of the fot(1) command and fotpx(1) command is as follows:

Command Command argument

fot[_-help] [_file1] [_file2]

fotpx


-help: Usage of fot(1) and fotpx(1) is output.

file1: The file1 is an input file. The default input file is the standard input file.

file2: The file2 is an output file. The default output file is the standard output file.

Examples are as follows:

1. File names in the command line are omitted (redirection)

$ fot <infile >outfile

2. File names in the command line are omitted (pipe)

$ command1 | fot | command2

3. File names in the command line are specified

$ fot infile outfile

4. Usage of fot(1) command is output

$ fot -help

B.2 Example of fot Command and fotpx CommandAn example of fot(1) command is as follows:

$ cat infile_aaaa

- 362 -

1bbbbcccc0dddd_eeee+ff$ fot infile | lprfot: invalid 1 line carriage control conventions in infile

_: Blank

Printing result:

aaaa Page 1bbbb Page 2cccc

ddddffee

B.3 Return Values of fot Command and fotpx CommandThe following table lists the return values set by the fot(1) command and fotpx(1) command.

Return value Status


1 fot(1) command or fotpx(1) command error

B.4 Diagnostic Messages of fot and fotpxThe fot command and fotpx command output diagnostic messages if the invalid file name is specified or the first character of the line ofthe input file is invalid.

Messages issued by fot command and fotpx command are in the following format:

fot: Message-text

fotpx: Message-text

- 363 -

Appendix C Endian Convert CommandThe endian convert commands fcvendian(1) and fcvendianpx(1) swap all bytes in type values within input file, and output the swappedvalues to the output file.

C.1 Command FormatThe format of the fcvendian(1) command and fcvendianpx(1) command is as follows:

Command Command argument

fcvendian_file1_file2_type

fcvendianpx


file1: An input file with values to swap.

file2: An output file to output the swapped values.

type: A type of values to swap (The value which can be specified is 2, 4, 8, or 16).

If no command arguments are specified, usage of fcvendian(1) and fcvendianpx(1) is output.

Examples are as follows:

Example1: To byteswap a file "in.data" with 8 byte values.

$ fcvendian in.data out.data 8

Example2: To output the usage of the fcvendian command.

$ fcvendian

C.2 Return Value of fcvendian(1) and fcvendianpx(1)The following table lists the return values set by the fcvendian(1) command and the fcvendianpx(1) command.

Return value Status


1 fcvendian(1) or fcvendianpx(1) command error

C.3 Diagnostic Messages of fcvendian(1) and fcvendianpx(1)If fcvendian(1) is not successful completion, the following diagnostics message is output to standard error file.

Usage: fcvendian infile outfile type

fcvendian: Invalid operand

fcvendian: Input file open error

fcvendian: Output file open error

fcvendian: Memory allocate error

If fcvendianpx(1) is not successful completion, the following diagnostics message is output to standard error file.

Usage: fcvendianpx infile outfile type

- 364 -

fcvendianpx: Invalid operand

fcvendianpx: Input file open error

fcvendianpx: Output file open error

fcvendianpx: Memory allocate error

C.4 NoteIf the file which the type of values is mixed in is swapped, the byteswap may be not successful completion.

- 365 -

Appendix D Job Operation SoftwareIn this appendix, the compilation and execution processes for using high-speed facility on Job Operation Software of Technical ComputingSuit. See the "Job Operation Software End-user's Guide" for details on Job Operation Software.

D.1 Management of the CPUs resourceThe CPUs resource that can be used is strictly managed on the job of Job Operation Software.

Number of CPUs that can be used on the job

This is the number of CPUs that can be used for a process on the job. See Section "D.1.1 Number of CPUs that can be used on theJob" for details.

Limits of Number of CPUs

The number of CPUs that can be used on the job is defined as the limit of the number of CPUs.

The number of threads when the program is executed is decided in consideration of the limit values of the number of CPUs. See Section"12.2.1.2.2 Number of Threads" for details of the number of the threads for the automatic parallelization. See "Number of Threads" inSection "12.3.1.2.3 Notes during Execution" for details of the number of threads for OpenMP.

D.1.1 Number of CPUs that can be used on the JobThe limit of the number of CPUs for one process is decided in consideration of CPUs resource and the number of processes in the virtualnode on the job. This value is defined as the number of CPUs that can be used on the job.

Details are shown as follows.

"The number of CPUs that can be used on job" = cpunum_on_node / procnum_on_node

cpunum_on_nodeThe number of CPUs that can be used on one virtual node.

1 to the number of all CPUs on one virtual node.

procnum_on_node

The number of processes on one virtual node.

For thread parallel processing jobs, it is 1.

For hybrid (process and thread) parallel processing jobs, it is 1 to the number ofall CPUs on one virtual node.

It is rounded down when it cannot be divided.

See the "Job Operation Software End-user's Guide" for details on job.

When MPI program is executed with VCOORD_FILE that is set the number of CPUs to a process, the value is the number of CPUs thatcan be used on the job. See the "MPI User's Guide" for details on MPI.

D.1.2 FLIB_USE_ALLCPU"The number of CPUs that can be used on the job" can be controlled using the environment variable FLIB_USE_ALLCPU. The valuethat can be set and the meaning are as follows. The default value is "FALSE".

TRUE

The number of CPUs that can be used on the job is cpunum_on_node. This means all CPUs thatcan be used on the virtual node is used regardless of the number of processes. Therefore, whenthe number of processes on the virtual node is two or more, these CPUs resource is shared amongthe processes. When the frequency calculated at the same time in each process is low etc. , theimprovement of the execution performance can be expected. In general use, the improvement ofthe execution performance cannot be expected.

You should use the following functions.

- -Kopenmp is used at compile time.

- FLIB_FASTOMP=FALSE is defined at run time.

- 366 -

- FLIB_SPINWAIT=0 is defined at run time.

When the above-mentioned is not used, the performance might be greatly downed like the deadlock.

See "Chapter 12 Multiprocessing" for details of -Kopenmp,FLIB_FASTOMP andFLIB_SPINWAIT.

When procnum_on_node is 1, "TRUE" and "FALSE" have the same effect.

When procnum_on_node is not 1, the on-chip hardware barrier and the sector-cache cannot beused.

See Section "D.1.1 Number of CPUs that can be used on the Job" for details aboutcpunum_on_node and procnum_on_node. See Section "D.3 On-chip Hardware Barrier" fordetails the on-chip hardware barrier, and Section "D.4 Sector-cache" for details about the sector-cache.

FALSEThe number of CPUs that can be used is Section "D.1.1 Number of CPUs that can be used on theJob".

When MPI program is executed with VCOORD_FILE that is set the number of CPUs to a process, the value is ignored. See the "MPIUser's Guide" for details on MPI.

D.1.3 FLIB_USE_CPURESOURCEThe environment variable FLIB_USE_CPURESOURCE controls the CPU resource manager on the job.

The allowed values their meanings are as follows. The default value is "TRUE".

TRUE The CPU resources on the job are managed.

FALSE The CPU resources on the job are not managed.

Therefore the number of CPUs is not managed, and you cannot use the following functions.

- D.2 CPU Binding

- D.3 On-chip Hardware Barrier

- D.4 Sector-cache

D.2 CPU BindingThe thread is bound to one CPU when the program using the automatic parallelization or OpenMP is executed on the job. CPUs used canbe strictly managed as described in Section "D.1 Management of the CPUs resource". Therefore the improvement of the performance canbe expected by CPU Binding.

See "Chapter 12 Multiprocessing for details of Multiprocessing.

D.2.1 FLIB_CPUBINDThe environment variable FLIB_CPUBIND controls the CPU Binding for thread. The following value can be set to FLIB_CPUBIND.The default value is "chip_pack".

chip_pack All threads are bound to same CPU Chip as much as possible.

chip_unpack Each thread is bound to a different CPU Chip as much as possible.

off Thread is not bound to CPU. If you control CPU Binding by yourself, this must be set.

When only one CPU Chip can be used, "chip_pack" and "chip_unpack" have the same effect.

When only one CPU on a CPU Chip can be used, "chip_pack" and "chip_unpack" have the same effect.

- 367 -

D.3 On-chip Hardware BarrierSince SPARC64VIIIfx, the SPARC64 processor has the on-chip hardware barrier. On the job, it can be used as the thread barrier. The on-chip hardware barrier is a hardware mechanism which facilitates high speed synchronization among threads in a CPU Chip, and improvesthe execution performance.

D.3.1 CompilationWhen generating the program that uses the on-chip hardware barrier as the thread barrier, the special compiler option is unnecessary. Itis automatically used when executing in the environment that can be used. The software barrier is used when executing in the environmentthat cannot be used.

The program generated using compiler option neither -Kparallel nor -Kopenmp doesn't use the thread barrier. For details about Execution,see Section "D.3.2 Execution.

D.3.2 ExecutionThe on-chip hardware barrier can be used only on the job. The on-chip hardware barrier is a hardware mechanism for synchronizationamong threads in a CPU Chip.

See the "Job Operation Software End-user's Guide" for detail of the job.

D.3.3 FLIB_CNTL_BARRIER_ERRWhen the on-chip hardware barrier cannot be used, the following messages are output, and the processing is continued using the softwarebarrier.

jwe1050i-w The hardware barrier couldn't be used and continues processing using the software barrier.

This diagnostic message (jwe1050i-w) error can be controlled using environment variable FLIB_CNTL_BARRIER_ERR.

The value that can be set and the meaning are as follows. Default is "TRUE".

TRUEThe diagnostic message (jwe1050i-w) error is detected. When the on-chip hardware barriercannot be used, this messages are output, and the processing is continued using the softwarebarrier.

FALSEThe diagnostic message (jwe1050i-w) error is not detected. When the on-chip hardwarebarrier cannot be used, the processing is continued using the software barrier.

D.3.4 FLIB_NOHARDBARRIERUsing the on-chip hardware barrier can be controlled using environment variable FLIB_NOHARDBARRIER. If the environment variableFLIB_NOHARDBARRIER exists, the on-chip hardware barrier cannot be used. When environment variable FLIB_NOHARDBARRIERis set (A set value is not necessary.), the software barrier is always used. Even if the on-chip hardware barrier can be used, it is not used.The diagnostic message (jwe1050i-w) error always occurs when environment variable FLIB_NOHARDBARRIER is set.

See Section "D.3.3 FLIB_CNTL_BARRIER_ERR" for details.

D.3.5 Notes- When the on-chip hardware barrier is used by the program made using compile option -Kopenmp, the environment variable

FLIB_FASTOMP shall not be FALSE. See Section "12.3 Parallelization by OpenMP Specifications" for the meaning and directionsof the environment variable FLIB_FASTOMP.

- When the on-chip hardware barrier is used, CPU Binding is needed. Therefore, when "off" is set for environment variableFLIB_CPUBIND, the on-chip hardware barrier cannot be used. See Section "D.2 CPU Binding" for details of CPU Binding.

- When the number of threads is 1, the on-chip hardware barrier cannot be used. See Section "12.2.1.2.2 Number of Threads" or Section"12.3.1.2.3 Notes during Execution" for the number of threads.

- 368 -

- When the number of CPUs to a process is three or less, a on-chip hardware barrier may not be able to use. The number of CPUs to aprocess should be set to four or more.

- When the job is a Swap Target Job and a Node-sharing Job, and the number of CPUs to a process is three or less, an on-chip hardwarebarrier cannot be used. The number of CPUs to a process should be set to four or more. See the "Job Operation Software End-user'sGuide" for details on job.

D.3.6 Usage example on the jobExample 1: A program that uses the on-chip hardware barrier is executed on a thread parallel processing job.

$ cat job.sh#!/bin/sh./a.out$ pjsub -L "node=1" job.sh

Example 2: A program that uses the on-chip hardware barrier is executed on a hybrid (process and thread) parallel processingjob (One job is executed in one dimension: The number of nodes is set to the number of processes.).

$ cat job.sh#!/bin/sh#PJM -L "node=12"#PJM --mpi "shape=12"#PJM --mpi "rank-map-bynode"mpiexec -n 12 a.out$ pjsub job.sh

See the "Job Operation Software End-user's Guide" for detail of the job.

D.4 Sector-cacheSince SPARC64VIIIfx, the SPARC64 processor has Sector-cache. On the job, the Sector-cache can be controlled.

See Section "9.17 Software Control of Sector Cache" about how to control the Sector-cache. Here, it explains only notes in execution.

D.4.1 Notes in executionWhen only one process is executed on a NUMA node, the control of the Sector-cache is possible.

See Section "D.1 Management of the CPUs resource" and the "Job Operation Software End-user's Guide" for details.

When it is not possible to execute by one process alone on a NUMA node and the Sector-cache control function is used, the diagnosticmessage jwe1047i-w is output as follows, the function is invalidated and processing is continued.

jwe1047i-w A sector cache couldn't be used.

On the following cases, Runtime Library cannot determine whether one process can reserve a NUMA node or not. Therefore behavior ofsector cache is indeterminate. Performance of program may be reduced, or program may be ended with diagnostic-message jwe1048i-u.

- Process is generated by functions other than Fujitsu MPI.

Example:

- FORTRAN FORK/SYSTEM/SH service function is used.

- C fork/system function is used.

- Threads are generated by functions other than Fujitsu Automatic Parallelization/OpenMP.

Example:

- Threads are controlled by pthread functions.

To invalidate the Sector-cache control function unconditionally regardless of the execution environment, "FALSE" is set to environmentvariable FLIB_SCCR_CNTL.

- 369 -

See Section "9.17.2.2 Software control with environment variables and optimization control rows" for details.

D.5 Large-Page FunctionThe large-page function enlarges the page size which is the unit of the memory management of hardware. If the page size is enlarged,TLBmiss reduces and the program runs faster.

The large-page function is enable for the executable program which is created specifying the -Klargepage option when linking. The stacksegment, data segment, heap segment and thread stack segment are enlarged for the executable program.

If the page size is enlarged, the used memory size may increase. If it increases the memory used by a command or a daemon, it may affectthe execution environment. The memory can decrease when a command or a daemon is created with specify the -Knolargepage option,and the effect of the execution environment can be decreased.

See "lpgparm(1)" for information of default size of large-page and how to change the large-page size.

- 370 -

Appendix E PreprocessorThis appendix explains the C language preprocessor and the Fortran preprocessor.

When a suffix of the source file is .F, .FOR, .F77, .F90, .F95 or .F03, or when the -Cpp option is specified, the preprocessor is executed.In this case, the -Ccpp or -Cfpp option can be specified to select the C language preprocessor or the Fortran preprocessor. The Fortranpreprocessor is selected by default.

For the details of the preprocessor options, see 2.2 Compiler Options.

E.1 C language preprocessorThe C language preprocessor preprocesses in conformance with C language specification.

E.2 Fortran preprocessorThe Fortran preprocessor gives more priority to the Fortran language specification than to the C language specification, and it preprocessesconsidering Fortran constructs, such as comment and continuation line indicator. Be careful to the followings:

- The Fortran constructs, such as continuation line indicator, comment, fixed and free source form and etc, are considered.

- The Fortran comment construct is also a target of macro substitution. However, the message is not output even if the error is in Fortrancomment construct.

- The Hollerith constant and H edit descriptor are not considered as character contexts, and are the targets of macro substitution.

- The B, O, X, Z, b, o, x, and z which represent binary, octal and hexadecimal constants are not affected by preprocessing directives,and are not the targets of macro substitution.

- Because the line number may be changed by the preprocessing of a Fortran source programs, be careful when this option is used. Inthis case, specify -P option and refer to the produced preprocessed output file.

- When the line exceeds limited column by the preprocessing directives and when -P option is specified, the control lines, such as"#line !fpp start" and "#line !fpp end", may be produced in the produced preprocessed output file. When the preprocessed output fileis specified as input file to the compiler, do not change those control lines because they are used by the compiler.

The following appendixes explain the features of the Fortran preprocessor with examples.

E.2.1 Continuation line of FortranThe continuation line of Fortran is considered.

- In fixed source form, column 6 indicates a continuation line. It is analyzed as the last character of the previous line continues to thecolumn 7 of the continuation line.

Example: aaa is substituted to bbb. #define aaa bbb !23456789 aa *a=1

- In free source form, the last character & and the first character & are analyzed as a character to indicate a continuation line. It isanalyzed as the character before & continues to the character after &.

Example: aaa is substituted to bbb. #define aaa bbb aa& &a=1

E.2.2 Comment line of FortranThe comment line of Fortran is considered.

- 371 -

- The comments of C language and Fortran can be specified. The comment of C language form is substituted to blanks.

Example: The comments of C language and Fortran are effective./* comment of C language comment of C language */ ! comment of Fortran

E.2.3 Comment of Fortran and comment of C language- When the beginning of C language comment "/*" is specified in the Fortran comment, the beginning of the C language comment is

not effective.

Example: The beginning of C language form comment is not effective. ! /* comment of Fortran Not comment of C language */

- When the comment of Fortran is specified in the C language comment, the comment of Fortran is not effective.

Example: Comment of Fortran is not effective and assignment statement is effective. /* ! */ kkk=1

E.2.4 Macro substitution in Fortran comment constructThe macro substitution in Fortran comment construct is considered.

- The macros in Fortran comment construct are the targets of substitution. It is the specification that considers that the OpenMP prefixand the optimization control line and etc begin with comment character.

Example: aaa in Fortran comment construct is substituted to bbb. #define aaa bbb !ocl noarraypad(aaa) integer::aaa(100) aaa=10

! aaa=20 !$omp parallel shared(aaa) :

- A message is not output even if the macro substitution error is detected in the Fortran comment construct.

Example: A message is not output even if the macro substitution error is detected in comment. #define a(i) b(2,i) a(5,10)=1 ! message is output. ! a(5,10)=1 message is not output.

E.2.5 Treatment of the characters specified after column 72 in fixed sourceform

In fixed source form, the characters specified after column 72 are ignored except comment and preprocessing directives. However, when-Fwide option is specified, columns 73 through 255 are effective.

Example: In fixed source form, the character C is specified at column 73, ABC is considered as AB, and ABC is not substituted to D. Specify -Fwide option in order to make columns 73 through 255 effective.#define ABC D! 1 2 3 4 5 6 7!23456789012345678901234567890123456789012345678901234567890123456789012

- 372 -

ABC * = 100 print *,D end

- 373 -

Appendix F Notes on Migration from FX10 System toFX100 System

This appendix provides notes on migrating from FX10 system (Generation Number:09 or later) to FX100 system.

For migrating from FX10 system (Generation Number:08 or earlier), refer to "Appendix G Compatibility Information (FX10 System)"also.

F.1 Error check in OpenMP standard at compilation timeThis note corresponds to the migration from FX10 system (Generation Number:11 or earlier).

Refer to "G.1.1 Error check in OpenMP standard at compilation time" about this note.

F.2 Error check in Fortran 2003 standard at compilation timeThis note corresponds to the migration from FX10 system (Generation Number:11 or earlier).

Refer to "G.1.2 Error check in Fortran 2003 standard at compilation time" about this note.

F.3 Enhancement intrinsic procedures and intrinsic moduleThis note corresponds to the migration from FX10 system (Generation Number:11 or earlier).

Refer to "H.1.3 Enhancement intrinsic procedures and intrinsic module" about this note.

F.4 The option which the compiler option -Kuxsimd needs ischanged

a. Changes

The option which the compiler option -Kuxsimd needs is changed.

[Previous version]

The compiler option -Kuxsimd needed only the compiler option -Ksimd.

[This version]

The compiler option -Kuxsimd needs the compiler option -KHPC_ACE in addition to the compiler option -Ksimd.

b. Influence

When the compiler option -KHPC_ACE is not specified, the compiler option -Kuxsimd is invalidated and the following messageis displayed.

frtpx: -Kuxsimd is invalidated because -KHPC_ACE is not specified.

c. Coping

Specify the compiler option -KHPC_ACE.

F.5 The function for canceling the program compilation that isforecasted to take a long time

This note corresponds to the migration from FX10 system (Generation Number:10 or earlier).

Refer to "G.2.1 The function for canceling the program compilation that is forecasted to take a long time" about this note.

- 374 -

F.6 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|openmp_noordered_reduction}


Refer to "G.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|openmp_noordered_reduction}" about this note.

F.7 Abolition of compilation options and the optimization controlspecifiers

1. -K{FLTLD|NOFLTLD} options and the optimization control specifiers {FLTLD|NOFLTLD}

[Note] The following description is changed only when the compiler option -KHPC_ACE2 is set.

a. Changes

The -K{FLTLD|NOFLTLD} options are abolished.

The optimization control specifiers {FLTLD|NOFLTLD} are abolished.

[Previous version]

The -K{FLTLD|NOFLTLD} options were effective.

The optimization control specifiers {FLTLD|NOFLTLD} were effective.

[This version]

The -K{FLTLD|NOFLTLD} options are invalid.

The optimization control specifiers {FLTLD|NOFLTLD} are invalid.

b. Influence

The -K{FLTLD|NOFLTLD} options become invalid. If these options are specified, the following message is output.

frtpx: -K[NO]FLTLD is invalidated because -KHPC_ACE is not specified.

And, the optimization control specifiers {FLTLD|NOFLTLD} become invalid. If these specifiers are specified, the followingmessage is output.

jwd1202i-i "filename", line number: Invalid optimization control specifier in optimization control line.

When the -K{FLTLD|NOFLTLD} options and the optimization control specifiers {FLTLD|NOFLTLD} become invalid, anormal load instruction is used.

c. Coping

Specify the -K{nf|nonf} options or the optimization control specifiers {NF|NONF}.

Although objects are compiled with the -KNOFLTLD option or the optimization control specifier NOFLTLD, it is possibleto execute the objects on SPARC64 XIfx by using non-faulting-load instruction.

2. -Kcpu option

a. Changes

The -Kcpu option is abolished.

[Previous version]

The -Kcpu option was effective.

- 375 -

[This version]

The -Kcpu option is invalid.

b. Influence

The -Kcpu option becomes invalid. If this option is specified, the following warning is output.

warning: Invalid suboption cpu specified for -K.

c. Coping

It is not necessary to action.

3. -K{prefetch_double_line_L2|prefetch_nodouble_line_L2} options and the optimization control specifiers{PREFETCH_DOUBLE_LINE_L2|PREFETCH_NODOUBLE_LINE_L2}

a. Changes

The -K{prefetch_double_line_L2|prefetch_nodouble_line_L2} options are abolished.

The optimization control specifiers {PREFETCH_DOUBLE_LINE_L2|PREFETCH_NODOUBLE_LINE_L2} areabolished.

[Previous version]

The -K{prefetch_double_line_L2|prefetch_nodouble_line_L2} options were effective.

The optimization control specifiers {PREFETCH_DOUBLE_LINE_L2|PREFETCH_NODOUBLE_LINE_L2} wereeffective.

[This version]

The -K{prefetch_double_line_L2|prefetch_nodouble_line_L2} options are invalid.

The optimization control specifiers {PREFETCH_DOUBLE_LINE_L2|PREFETCH_NODOUBLE_LINE_L2} areinvalid.

b. Influence

The -K{prefetch_double_line_L2|prefetch_nodouble_line_L2} options become invalid. If these options are specified, thefollowing warning is output.

warning: Invalid suboption prefetch_[no]double_line_L2 specified for -K.

And, the optimization control specifiers {PREFETCH_DOUBLE_LINE_L2|PREFETCH_NODOUBLE_LINE_L2}become invalid. If these specifiers are specified, the following warning is output.

jwd1202i-i "filename", line number: Invalid optimization control specifier in optimization control line.

c. Coping


4. -K{simd_region_constant|nosimd_region_constant} options

a. Changes

This note corresponds to the migration from FX10 system (Generation Number:07 or later).

The -K{simd_region_constant|nosimd_region_constant} options are abolished.

[Previous version]

The -K{simd_region_constant|nosimd_region_constant} options were effective.

[This version]

The -K{simd_region_constant|nosimd_region_constant} options are invalid.

- 376 -

b. Influence

The -K{simd_region_constant|nosimd_region_constant} options become invalid. If these options are specified, the followingwarning is output.

warning: Invalid suboption [no]simd_region_constant specified for -K.

c. Coping


F.8 The -Komitfp option is induced by -Kfast, and maintenance ofthe trace back information


Refer to "G.3.1 The -Komitfp option is induced by -Kfast, and maintenance of the trace back information" about this note.

F.9 Change of value of macro along with the compiler version upThis note corresponds to the migration from FX10 system (Generation Number:09 or earlier).

Refer to "G.3.2 Change of value of macro along with the compiler version up" about this note.

F.10 Change of the name of temporary object file and the directoryto store


Refer to "G.3.3 Change of the name of temporary object file and the directory to store" about this note.

F.11 Change the default value of the -Ksimd[=level] optiona. Changes

The default value of the -Ksimd[=level] option is changed.

[Previous version]

1. If a value was not specified for level of the -Ksimd[=level] option, -Ksimd=1 was set.

2. When the -O2 option or higher was set, the -Ksimd=1 option was induced. (*1)

3. If a value was not specified for level of the -Ksimd[=level] option, the -Q or -Nlst option output -Ksimd=1.

[This version]

1. If a value is not specified for level of the -Ksimd[=level] option, -Ksimd=auto is set.

2. When the -O2 option or higher is set, the -Ksimd=auto option is induced. (*1)

3. If a value is not specified for level of the -Ksimd[=level] option, the -Q or -Nlst option output -Ksimd=auto.

*1) When the -Kfast option is specified, the -O3 option is set. When the -O[n] option is not specified, the -O2 option is set.

b. Influence

SIMD extensions are promoted for loops that contain IF constructs in the following cases:

- The level value of the -Ksimd[=level] option is not specified. And,

- The -O2 option or higher is set.

c. Coping

To keep the SIMD condition the same as the previous version, specify the -Ksimd=1 option.

- 377 -

F.12 Specification change of the optimization control specifiersSIMD and UXSIMD

[Note] The following description is changed only when the compiler option -KHPC_ACE2 is set.

a. Changes

Specifications of the SIMD and UXSIMD specifiers are changed.

Hardware restrictions of data alignment for SIMD store instructions on FX100 system are relaxed.

Therefore, SIMD store instructions are usable without the ALIGNED or UNALIGNED specifier which is following the SIMD andUXSIMD specifiers.

[Previous version]

Previous specifications of the SIMD and UXSIMD specifiers were as follows.

!OCL SIMD[(ALIGNED | UNALIGNED)]

!OCL UXSIMD[(ALIGNED | UNALIGNED)]

[This version]

Specifications of the SIMD and UXSIMD specifiers are as follows.

!OCL SIMD

!OCL UXSIMD

b. Influence

The compiler ignores ALIGNED and UNALIGNED specifiers following the SIMD and UXSIMD specifiers.

c. Coping


- 378 -

Appendix G Compatibility Information (FX10 System)This appendix provides compatibility information as notes on migrating.

G.1 Migrating to V2.0L20(Generation Number:12)

G.1.1 Error check in OpenMP standard at compilation timea. Changes

The error messages are output at compilation time for incorrect program in OpenMP standard.

- Diagnostic message jwd1838i-s is output at compilation time under the following conditions:

1. -Kopenmp option is specified. And,

2. A common block name appears in SAVE statement. And,

3. The common block appears in THREADPRIVATE directive. And,

4. The common block does not appear in COMMON statement.



2. A common block is declared in COMMON statement in a module. And,

3. The common block appears in THREADPRIVATE directive. And,

4. The module name appears in USE statement in another program unit. And,

5. In that program unit, the common block appears in COPYIN or COPYPRIVATE clause. And,

6. In that program unit, the common block name does not appear in COMMON statement.



2. An automatic object is declared in a subprogram. And,

3. The variable in specification expression of the automatic object appears in THREADPRIVATE directive.

[Previous version]

Diagnostic messages jwd1838i-s, jwd1837i-s, or jwd2038i-s was not output at compilation time and object was created.

[This version]

Diagnostic messages jwd1838i-s, jwd1837i-s, or jwd2038i-s is output at compilation time.

b. Influence

Diagnostic messages jwd1838i-s, jwd1837i-s, or jwd2038i-s is output at compilation time and object is not created.

c. Coping

Modify program as following:

- Make common blocks in COPYIN clause, COPYPRIVATE clause, or THREADPRIVATE directive, specified in COMMONstatement in same scoping unit. Or,

- Do not reference local variable in specification expression.

G.1.2 Error check in Fortran 2003 standard at compilation timea. Changes

The error messages are output at compilation time for incorrect program in the Fortran 2003 standard.

- 379 -


1. Procedure declaration statement appears. And,

2. Declaration type specifier in interface of the procedure declaration statement. And,

3. Letter or digit appears after the declaration type specifier.


1. Procedure declaration statement appears. And,

2. CHARACTER intrinsic type specifier appears in interface of the procedure declaration statement. And,

3. KIND= appears after the type specifier. And,

4. LEN= does not appear after the type specifier. And,

5. Right parenthesis does not appear after the type specifier.

- Diagnostic message jwd1176i-s or jwd1343i-s is output at compilation time under the following conditions:

1. Assumed size array is declared. And,

2. Lower bound is omitted and :* appears in last dimension of the assumed size array.


1. SELECT TYPE construct or ASSOCIATE construct appears. And,

2. Brach appears from out of the scope to the inside the scope. And,

3. Diagnostic message jwd1065i-w is output.

- Diagnostic message jwd1272-s is output at compilation time under the following conditions:

1. Interface block appears in module program. And,

2. USE statement specified module of 1 appears in program unit. And,

3. Interface appears in program unit of 2. And,

4. Generic specifier appears in the interface of 3. And,

5. PROCEDURE statement appears in interface body of 3. And,

6. The procedure of the interface body of 1 appears two times or more in interface body of 3. And,

7. The generic specifier does not refer.

[Previous version]

Diagnostic message jwd1118i-s, jwd1176i-s, jwd1272i-s, jwd1343i-s, or jwd1954i-s was not output at compilation time andobject was created.

[This version]

Diagnostic message jwd1118i-s, jwd1176i-s, jwd1272i-s, jwd1343i-s, or jwd1954i-s is output at compilation time.

b. Influence

Diagnostic message jwd1118i-s, jwd1176i-s, jwd1272i-s, jwd1343i-s, or jwd1954i-s is output at compilation time and object is notcreated.

c. Coping

Modify program as following:

- Fix the procedure declaration statement.

- Specify lower bound value or modify :* to * in last dimension of assumed size array.

- Do not specify branch to SELECT TYPE construct or ASSOCIATE construct from outside.

- Do not specify same procedure name two times or more in interface of generic specifier.

- 380 -


G.2.1 The function for canceling the program compilation that is forecastedto take a long time

a. Changes

The compilation is canceled when the compiler forecasts that it takes a long time (24 hours or more as a guide) to compile theprogram.

[Previous version]

The compilation was not canceled when it took a long time to compile.

[This version]

The compilation is canceled after outputting the following message when the compiler forecasts that it takes a long time tocompile.

jwd8695i-u The compilation was canceled because it was forecasted to take a long time.(name:proc-name)

proc-name: Procedure name

b. Influence

It may not be able to compile the program which was able to compile previously.

c. Coping

Specify the -Nnocancel_overtime_compilation option at compilation.

G.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|openmp_noordered_reduction}

a. Changes

The compiler option -K{ ordered_omp_reduction | noordered_omp_reduction } is changed to -K{ openmp_ordered_reduction |openmp_noordered_reduction }.

The option which is output by the compiler option -Q or -Nlst is changed from -K{ ordered_omp_reduction |noordered_omp_reduction } to -K{ openmp_ordered_reduction | openmp_noordered_reduction }.

[Previous version]

The -Q or -Nlst option output -K{ ordered_omp_reduction | noordered_omp_reduction }.

[This version]

The -Q or -Nlst option outputs -K{ openmp_ordered_reduction | openmp_noordered_reduction }.

b. Influence

When -K{ ordered_omp_reduction | noordered_omp_reduction } option is specified, the following warning message is output and-K{ openmp_ordered_reduction | openmp_noordered_reduction } option is effective.

warning: obsolete option -Kordered_omp_reduction changed -Kopenmp_ordered_reduction.

c. Coping

It is not necessary to action. It is recommended to specify the -K{ openmp_ordered_reduction | openmp_noordered_reduction }option.

- 381 -


G.3.1 The -Komitfp option is induced by -Kfast, and maintenance of thetrace back information

a. Changes

The -Komitfp option is induced by -Kfast.

When the -Komitfp option is set, the trace back information is not kept.

[Previous version]

The -Komitfp option was not prepared.

[This version]

The -Komitfp option is added.

The -Komitfp option is induced by -Kfast.


b. Influence


c. Coping

Specify the -Knoomitfp option to maintain the trace back information, and re-compile.

G.3.2 Change of value of macro along with the compiler version upa. Changes

The value of macro __frt_version is changed along with the compiler version up.

[Previous version]

The value of the macro __frt_version was 600.

[This version]

The value of the macro __frt_version is 700.

b. Influence

When the value of macro __frt_version is used in the program, the behavior of the program is different from former.

c. Coping

Correct the program to correspond to the value after the change.

G.3.3 Change of the name of temporary object file and the directory to storea. Changes

The name of temporary object file and the directory to store is changed.

[Previous version]

The name of temporary object file was "input-filename.o" and the directory to store was current directory.

[This version]

The name of temporary object file is unique filename and the directory to store is the temporary directory.

The temporary directory is a directory specified by the environment variable TMPDIR.

When the environment variable TMPDIR is not specified, /tmp is used instead of TMPDIR.

- 382 -

b. Influence

- The capacity of the temporary directory increases temporarily: the temporary object file is removed automatically.

- When two or more input files were specified, their object files had been left in current directory, but now they are not left.

c. Coping

When the capacity of a temporary directory is insufficient, delete unnecessary files.

Specify the compiler option -c to leave object file in current directory.


G.4.1 Change of values of macro and named constant along with thesupport of OpenMP API version 3.1 specifications

a. Changes

The values of the following macro and named constant are changed along with the support of the OpenMP API version 3.1specifications.

- The value of macro _OPENMP is changed.

- The values of named constant openmp_version defined by include file omp_lib.h and module omp_lib are changed.

[Previous version]

- When compile option -Kopenmp was specified, -D_OPENMP=200805 was effective.

- The value of named constant openmp_version was 200805.

[This version]

- When compile option -Kopenmp is specified, -D_OPENMP=201107 is effective.

- The value of named constant openmp_version is 201107.

b. Influence

When the value of macro _OPENMP or named constant openmp_version is used in the program, the behavior of the program isdifferent from former.

c. Coping

Correct the program to correspond to the value after the change.

G.4.2 Enhancement Fortran 2008 intrinsic procedures and intrinsicmodules

a. Changes

1. The procedure name IS_CONTIGUOUS without the EXTERNAL attribute become intrinsic procedure.

2. When intrinsic module ISO_FORTRAN_ENV is referred, COMPILER_OPTIONS and COMPILER_VERSION becomeintrinsic module procedures by ISO_FORTRAN_ENV.When intrinsic module ISO_C_BINDING is referred, C_SIZEOF becomes intrinsic module procedure by ISO_C_BINDING.

3. When intrinsic module ISO_FORTRAN_ENV is referred, the following names become named constants defined byISO_FORTRAN_ENV.

- CHARACTER_KINDS

- INT8, INT16, INT32, INT64

- REAL32, REAL64, REAL128

- INTEGER_KINDS

- 383 -

- LOGICAL_KINDS

- REAL_KINDS

[Previous version]

1. The procedure name IS_CONTIGUOUS was not intrinsic procedure name.

2. Neither COMPILER_OPTIONS, COMPILER_VERSION nor C_SIZEOF were procedures of the intrinsic module.

3. The following names were not defined by the intrinsic module.

- CHARACTER_KINDS



- INTEGER_KINDS

- LOGICAL_KINDS

- REAL_KINDS

[This version]

1. The procedure name IS_CONTIGUOUS become intrinsic procedure.

2. COMPILER_OPTIONS, COMPILER_VERSION, and C_SIZEOF become procedures of intrinsic moduleISO_FORTRAN_ENV or ISO_C_BINDING.

3. The following names become named constants of intrinsic module ISO_C_FORTRAN_ENV.

- CHARACTER_KINDS

- INT8,INT16,INT32,INT64

- REAL32,REAL64,REAL128

- INTEGER_KINDS

- LOGICAL_KINDS

- REAL_KINDS

b. Influence

1. Diagnostic message is output at compilation time when characteristics of procedure (function, subroutine, type of functionresult, number of argument or type of argument) and IS_CONTIGUOUS are different.Program is abnormally ended or result is an incorrect at execution time when all of the characteristics for the procedure arecorresponding.

2. The diagnostic message is output at compilation time when the procedure characteristics (function, subroutine, type of thefunction result, and number of arguments) of COMPILER_OPTIONS, COMPILER_VERSION, or C_SIZEOF are notcorresponding, and referring to intrinsic module ISO_FORTRAN_ENV or ISO_C_BINDING.Program is abnormally ended or result is an incorrect at execution time when all of the characteristics for the procedure arecorresponding.

3. The diagnostic message is output at compilation time when the same user definition name exists as named constants of theadditional, and referring to intrinsic module ISO_FORTRAN_ENV.

c. Coping

1. Correct the source program to specify the EXTERNAL attribute for user function IS_CONTIGUOUS.

2. The use association names only specify by ONLY option of intrinsic module ISO_FORTRAN_ENV or ISO_C_BINDING,and must not refer COMPILER_OPTIONS, COMPILER_VERSION, or C_SIZEOF.

3. The use association names only specify by ONLY option of intrinsic module ISO_FORTRAN_ENV, and must not refer thefollowing names.

- CHARACTER_KINDS

- 384 -



- INTEGER_KINDS

- LOGICAL_KINDS

- REAL_KINDS

G.4.3 Bound remapping list and pointer target of two rank or more in pointerassignment statement

a. Changes

It corresponds to the change in the Fortran 2008 standard.

Diagnostic message jwd2348i-s is output at compilation time under the following conditions:

- A bound remapping list is declared in pointer assignment statement. And,

- The rank of pointer target is two or more. And,

- The pointer target is not simply CONTIGUOUS.

[Previous version]

Diagnostic message jwd2348i-s was not output at compilation time and object was created.

[This version]

Diagnostic message jwd2348i-s is output at compilation time.

b. Influence

Diagnostic message jwd2348i-s is output at compilation time and object is not created.

c. Coping

Modify pointer target with simply CONTIGUOUS.

G.4.4 Non pure final subroutine referred from pure procedurea. Changes


Diagnostic message jwd2597i-s is output at compilation time under the following conditions:

- There is a non pure final subroutine referred from pure procedure. And,

- The final subroutine is called by local variable.

[Previous version]


[This version]


b. Influence


c. Coping

Fix final subroutine to pure.

- 385 -

G.4.5 Array argument that is not CONTIGUOUS in intrinsic module functionC_LOC

a. Changes


Diagnostic message jwd2351i-s is output at compilation time under the following condition:

- An array without CONTIGUOUS appears in intrinsic module function C_LOC.

[Previous version]


[This version]


b. Influence


c. Coping

Fix to specify CONTIGUOUS array in intrinsic module function C_LOC.

G.4.6 Runtime message (jwe1007i-s) changea. Changes

The runtime message (jwe1007i-s) is changed.

[Previous version]

The value specified for a type parameter must be equal to the value specified in the declaration of the allocatable variable.

[This version]

The value of type parameter of the allocate object must not differ from the value of corresponding type parameter of sourceexpression or type specification.

b. Influence

The changed message is output.

When errmsg-variable is specified for ERRMSG, and the error of jwe1007i-s occurs in the ALLOCATE statement, the messagecontent of the variable is changed.

c. Coping



G.5.1 Error message output when specifying parallelization options areomitted at linking

1. -Kparallel option

a. Changes

The linking process when -Kparallel option is not specified for the automatically parallelized object at linking is changed.

[Previous version]

An executable file using serial processing was created without the error message.

- 386 -

[This version]

The following error message is output, and the linking process is suspended.

command: fatal: -Kparallel option is not specified at linking of object files to which -Kparallel applied at compiling. The linking process is suspended.

command: { frt | frtpx }

b. Influence

An executable file is not created if -Kparallel option is not specified for the automatically parallelized object at linking.

c. Coping

Specify -Kparallel option at linking if that error message is output.

d. Note

The target files for checking whether specifying -Kparallel option is omitted are object files (*.o) and libraries (*.a, *.so)input for compile command. Libraries specified by -l option are not target.

2. -Kopenmp option

a. Changes

The linking process when -Kopenmp option is not specified for the object parallelized by OpenMP at linking is changed.

[Previous version]

An executable file using serial processing was created without the error message.

[This version]

The following error message is output, and the linking process is suspended.

command: fatal: -Kopenmp option is not specified at linking of object files to which -Kopenmp applied at compiling. The linking process is suspended.

command: { frt | frtpx }

b. Influence

An executable file is not created if -Kopenmp option is not specified for the parallelized object by OpenMP at linking.

c. Coping

Specify -Kopenmp option at linking if that error message is output.

d. Note

The target files for checking whether specifying -Kopenmp option is omitted are object files (*.o) and libraries (*.a, *.so)input for compile command. Libraries specified by -l option are not target.


G.6.1 Message output for unrecognized compiler optiona. Changes

Warning message is output when unrecognized compiler option is specified.

[Previous version]

No warning message was output when unrecognized compiler option was specified.

[This version]


- 387 -

b. Influence


c. Coping

Specify recognizable compiler option. Use the -W option to specify linker option as follows:

-Wl,linker-option (*1)

(*1) "l" is lowercase L.

G.7 Migrating to V1.0L20

G.7.1 Runtime message (jwe0220i-e) changea. Changes

The runtime message (jwe0220i-e) is changed.

[Previous version]

In ATAN2(x1,x2), x1.eq.0.0 .and. x2.eq.0.0.

[This version]

In ATAN(x1,x2) or ATAN2(x1,x2), x1.eq.0.0 .and. x2.eq.0.0.

b. Influence

The changed message is output.

c. Coping


G.7.2 Changes to the specification when SOURCE= specifier appears andan allocatable component is "allocated" in ALLOCATE statement

a. Changes

It corresponds to the change in the Fortran standard.

If SOURCE= specifier appears and the allocation status of allocatable component is "allocated" in ALLOCATE statement, theallocation status for corresponding allocatable component is changed "allocated" from "unallocated".

[Previous version]

Any allocatable components had an allocation status of "unallocated" in ALLOCATE statement.

[This version]

If SOURCE= specifier appears and the allocation status of allocatable component is "allocated" in ALLOCATE statement, theallocation status for corresponding allocatable component is "allocated".

Any allocatable components have an allocation status of "unallocated" unless the conditions.

b. Influence

If SOURCE= specifier appears and the allocation status of allocatable component is" allocated" in ALLOCATE statement, theallocation status for corresponding allocatable component is "allocated".

Therefore, if the component is allocated after the execution, the runtime message jwe1001i-s is output at execution time.

It is because of becoming a reallocation status.

- 388 -

c. Coping

If SOURCE= specifier appears and the allocation status of allocatable component is" allocated" in ALLOCATE statement, theallocation status for corresponding allocatable component is "allocated". Therefore, if the component is allocated after the execution,modify either of the following program correction.

- Change the allocation status for corresponding allocatable component to "unallocated" from "allocated" before the executionof SOURCE= specifier appeared allocatable statement, or

- Change the allocation status for corresponding allocatable component to "unallocated" from "allocated" after the execution ofSOURCE= specifier appeared allocatable statement.

G.7.3 Support Fortran 2008 intrinsic proceduresa. Changes

The following procedure names without the EXTERNAL attribute become intrinsic procedure.

ACOSH,ASINH,ATANH,BGE,BGT,BLE,BLT,DSHIFTL,DSHIFTR,HYPOT,LEADZ,MASKL,MASKR,MERGE_BITS,POPCNT,POPPAR,SHIFTA,SHIFTL,SHIFTR,STORAGE_SIZE,TRAILZ

[Previous version]

The procedure names became user defined external procedure.

[This version]

The procedure names become intrinsic procedure.

b. Influence

Diagnostic message is output at compilation time when characteristics (function, subroutine, type of function result, number ofargument or type of argument) of procedure are different.

Program ends terminate abnormally or result is an incorrect at execution time when all of the characteristics for the procedure arecorresponding.

c. Coping

Specify an EXTERNAL attribute in source program.

G.7.4 Change when specifying the -O option before specifying the -g optiona. Changes

The -O option is effective when specifying the option which makes the -O option effective before specifying the -g option.

[Previous version]

The -g option set the -O0 option.

To make the -O option effect, the -O option had to be specified after the -g option.

Example 1: -g and -O0 options were effective

$ frtpx -O3 -g a.f90

Example 2: -g and -O3 options were effective

$ frtpx -g -O3 a.f90

[This version]

When the -g option is specified, and the option which makes the -O1 or higher option effective is not specified, the -O0 optionis effective.

Example 1: -g and -O0 options are effective

$ frtpx -g a.f90

- 389 -

Example 2: -g and -O3 options are effective

$ frtpx -O3 -g a.f90

$ frtpx -g -O3 a.f90

b. Influence

When -O1 or higher option is specified before the -g option is specified, the optimization becomes effective, compilation time mayincrease, and the size of object program may increase.

c. Coping

Specify the -O0 option to suppress the optimization.

- 390 -

Appendix H Compatibility Information (FX100 System)This appendix provides compatibility information as notes on migrating.

H.1 Migrating to V2.0L20(Generation Number:03)

H.1.1 Error check in OpenMP standard at compilation timeRefer to "G.1.1 Error check in OpenMP standard at compilation time" about this note.

H.1.2 Error check in Fortran 2003 standard at compilation timeRefer to "G.1.2 Error check in Fortran 2003 standard at compilation time" about this note.

H.1.3 Enhancement intrinsic procedures and intrinsic modulea. Changes

1. The following procedure names without EXTERNAL attribute become intrinsic procedures:

- ATOMIC_DEFINE

- ATOMIC_REF

- CO_MAX

- CO_MIN

- CO_SUM

- IMAGE_INDEX

- LCOBOUND

- NUM_IMAGES

- THIS_IMAGE

- UCOBOUND

2. When intrinsic module ISO_FORTRAN_ENV is referred, the following names become named constants defined byISO_FORTRAN_ENV:

- ATOMIC_INT_KIND

- ATOMIC_LOGICAL_KIND

- STAT_STOPPED_IMAGE

- STAT_LOCKED

- STAT_LOCKED_OTHER_IMAGE

- STAT_UNLOCKED

3. When intrinsic module ISO_FORTRAN_ENV is referred, a LOCK_TYPE become derived type name defined byISO_FORTRAN_ENV.

[Previous version]

1. The following procedure names without EXTERNAL attribute were not intrinsic procedures:

- ATOMIC_DEFINE

- ATOMIC_REF

- CO_MAX

- 391 -

- CO_MIN

- CO_SUM

- IMAGE_INDEX

- LCOBOUND

- NUM_IMAGES

- THIS_IMAGE

- UCOBOUND

2. The following names were not defined by the intrinsic module ISO_FORTRAN_ENV:

- ATOMIC_INT_KIND



- STAT_LOCKED


- STAT_UNLOCKED

- LOCK_TYPE

[This version]

1. The following procedure names without EXTERNAL attribute become intrinsic procedures.

- ATOMIC_DEFINE

- ATOMIC_REF

- CO_MAX

- CO_MIN

- CO_SUM

- IMAGE_INDEX

- LCOBOUND

- NUM_IMAGES

- THIS_IMAGE

- UCOBOUND

2. When intrinsic module ISO_FORTRAN_ENV is referred, the following names become named constants defined byISO_FORTRAN_ENV.

- ATOMIC_INT_KIND



- STAT_LOCKED


- STAT_UNLOCKED

3. When intrinsic module ISO_FORTRAN_ENV is referred, a LOCK_TYPE become derived type name defined byISO_FORTRAN_ENV.

- 392 -

b. Influence

1. Diagnostic message is output at compilation time when characteristics of procedures (function, subroutine, type of functionresult, number of argument, or type of argument) are different.Program is abnormally ended or result becomes incorrect at execution time when all of the characteristics of the proceduresare same.

2. The diagnostic message is output at compilation time when the same user-defined name exists as named constants of theadditional, and referring to intrinsic module ISO_FORTRAN_ENV.

c. Coping

1. Correct the source program to specify the EXTERNAL attribute for user function which has same name as above functions.

2. Use ONLY option for USE statement of ISO_FORTRAN_ENV to make above named constants not to be use associated.

H.1.4 Change of the compilation message when the compiler option -K{FLTLD|NOFLTLD} is specified

a. Changes

Compilation message is changed when the compiler option -K{FLTLD|NOFLTLD} is specified.

[Previous version]

The following warning message was output when the compiler option -K{FLTLD|NOFLTLD} was specified.

frtpx: warning: Invalid suboption [NO]FLTLD specified for -K.

[This version]

The following message is output when the compiler option -KHPC_ACE2 is valid and the compiler option -K{FLTLD|NOFLTLD} is specified.

frtpx: -K[NO]FLTLD is invalidated because -KHPC_ACE is not specified.

b. Influence

The message of [This version] is output when the compiler option -KHPC_ACE2 is valid and the compiler option -K{FLTLD|NOFLTLD} is specified.

c. Coping


H.2 Migrating to V2.0L10(Generation Number:02)

H.2.1 The function for canceling the program compilation that is forecastedto take a long time

Refer to "G.2.1 The function for canceling the program compilation that is forecasted to take a long time" about this note.

H.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|openmp_noordered_reduction}

Refer to "G.2.2 The compiler option -K{ordered_omp_reduction|noordered_omp_reduction} is changed to -K{openmp_ordered_reduction|openmp_noordered_reduction}" about this note.

- 393 -

Index[Special characters]

-#...............................................................................................57-###...........................................................................................57$ edit descriptor............................................................... 116,137&END.....................................................................................120&namelist name......................................................................120

[A]-A.............................................................................................. 17-a............................................................................................... 60-AA...................................................................................... 17,92ACTION= Specifier........................................................ 124,163-Ad....................................................................................... 17,93-AE............................................................................................17-Ae............................................................................................ 17-Ai.............................................................................................17ALLOCATE Statement.......................................................... 108-Ap............................................................................................ 18-Aq....................................................................................... 18,93arithmetic IF statement........................................................... 105array declaration..................................................................... 360array element..................................................................... 25,206array section.........................................................25,206,359,361assembler.................................................................................... 4assigned GO TO statement.............................................. 105,229assignment statements with overlapping character positions.104ASSOCIATE construct...........................................................359-AT............................................................................................17-AU........................................................................................... 17-Aw........................................................................................... 18-Ay............................................................................................ 18-Az............................................................................................ 18

[B]BACKSPACE statement........................................................ 154big endian data........................................................................161BLOCKSIZE= Specifier........................................................ 125

[C]-c............................................................................................... 14-C......................................................................................... 18,63CARRIAGECONTROL= Specifier....................................... 125CASE construct...................................................................... 359-Cca4a8.....................................................................................22-Ccd4d8.................................................................................... 21-CcdDR16.................................................................................22-CcdRR8................................................................................... 20-CcI4I8......................................................................................19-CcL4L8....................................................................................20-CcR4R8................................................................................... 21-CcR8R16................................................................................. 23-CcdII8......................................................................................19-CcdLL8....................................................................................20-Cfpp.........................................................................................19character type data.................................................................... 92-Cl........................................................................................ 18,94

command.................................................................................... 4common block..................................................................... 25,92COMMON statement............................................................... 99compiler option...........................................................................5complex type data.....................................................................89control information................................................................. 134CONVERT= Specifier............................................................125-Cpp..................................................................................... 18,19cpp command..............................................................................4

[D]-D.............................................................................................. 23-d...............................................................................................61DEALLOCATE Statement.....................................................108debugging............................................................................... 170debugging check function.......................................................202debugging functions............................................................... 202default implicit typing.............................................................. 17diagnostic message................................................................... 14direct access input/output statement....................................... 137DO construct...........................................................................359DO statement.......................................................................... 106

[E]-E.............................................................................................. 24-e............................................................................................... 61-Ec......................................................................................24,103-Eg............................................................................................ 24endfile record..........................................................................120ENDFILE statement............................................................... 155EQUIVALENCE statement....................................................100error........................................................................................ 170error condition........................................................................ 136error level..................................................................................70error monitor...........................................................................170error processing...................................................................... 170ERROR STOP statement........................................................107executable program...................................................................14executable program output....................................................... 86execution command..................................................................60execution command environment variable...............................65execution command format...................................................... 60execution command return value..............................................66expression............................................................................... 102

[F]-f................................................................................................14-fi.............................................................................................. 14FILE= Specifier...................................................................... 122file connection........................................................................ 112file positioning statement........................................................154-Fixed........................................................................................24FLIB_CPU_AFFINITY...................................................315,349FLIB_FASTOMP................................................................... 346FLIB_SPINWAIT.................................................................. 346FLIB_USE_STOPCODE......................................................... 68

- 394 -

-fmsg_num................................................................................14FORT90CPX............................................................................ 58FORT90L..................................................................................65Fortran compiler......................................................................... 1Fortran library.............................................................................1fot............................................................................................362fotpx........................................................................................362fpp command.............................................................................. 4-Free..........................................................................................24frtpx............................................................................................ 1-fs.............................................................................................. 14function...................................................................................359-fw.............................................................................................14-Fwide.......................................................................................24

[G]-g..........................................................................................14,61-G.............................................................................................. 63GO TO statement....................................................................105

[H]-H.............................................................................................. 24-Ha..................................................................................... 24,204-He..................................................................................... 24,208-Hf.............................................................................................25-Ho............................................................................................ 25-Hs............................................................................................ 25-Hu..................................................................................... 25,206-Hx..................................................................................... 25,208

[I]-I................................................................................................25-i................................................................................................61IEEE-format floating-point data.............................................158IF construct............................................................................. 359implied DO-loop.....................................................................359INCLUDE file.......................................................................... 25INCLUDE line..........................................................................25INQUIRE statement parameter.............................................. 128integer type data........................................................................89internal file..............................................................................117internal file input/output statement.........................................150intrinsic function names with the type declared..................... 111IOSTAT= Specifier................................................................ 135

[J]Job Oparation Software.......................................................... 366

[K]-K.............................................................................................. 26-Kadr44.....................................................................................26-Kadr64.....................................................................................26-Karraypad_const......................................................................26-Karraypad_expr.......................................................................26-Karray_merge..........................................................................27-Karray_merge_common..........................................................27-Karray_merge_local................................................................ 27-Karray_merge_local_size........................................................27

-Karray_subscript..................................................................... 27-Karray_subscript_element.......................................................27-Karray_subscript_elementlast................................................. 27-Karray_subscript_rank............................................................ 27-Karray_subscript_variable...................................................... 28-Karray_transform.................................................................... 28-Kauto........................................................................... 27,28,343-Kautoobjstack..........................................................................28-Kcalleralloc............................................................................. 28-Kcommonpad................................................................... 29,235-Kdalign.................................................................................... 29-Kdynamic_iteration.................................................................29-Keval................................................................................ 29,224-Kfast........................................................................................ 30-Kfed.........................................................................................30-Kfenv_access...........................................................................30-KFLTLD..................................................................................30-Kfp_contract............................................................................31-Kfp_relaxed.............................................................................31-Kfsimple.................................................................................. 31-KHPC_ACE............................................................................ 31-KHPC_ACE2.......................................................................... 31-Kilfunc.....................................................................................32-Kindependent.......................................................................... 32-Kinstance.................................................................................32-Kintentopt................................................................................32-Klargepage.............................................................................. 33-Kloop_blocking.......................................................................33-Kloop_fission.......................................................................... 33-Kloop_fission_if......................................................................33-Kloop_fusion...........................................................................33-Kloop_interchange.................................................................. 34-Kloop_noblocking...................................................................33-Kloop_nofission...................................................................... 33-Kloop_nofission_if..................................................................33-Kloop_nofusion.......................................................................34-Kloop_nointerchange.............................................................. 34-Kloop_nopart_parallel.............................................................34-Kloop_nopart_simd.................................................................34-K loop_noversioning............................................................... 34-Kloop_part_parallel.................................................................34-Kloop_part_simd.....................................................................34-Kloop_versioning.................................................................... 34-Klto..........................................................................................34-Kmfunc....................................................................................35-Knf...........................................................................................36-Knoalias...................................................................................26-Knoauto............................................................................ 28,343-Knoautoobjstack......................................................................28-Knocalleralloc......................................................................... 29-Knodalign................................................................................ 29-Knodynamic_iteration.............................................................29-Knoeval................................................................................... 30-Knofed.....................................................................................30-Knofenv_access.......................................................................30-KNOFLTLD............................................................................30-Knofp_contract........................................................................31

- 395 -

-Knofp_relaxed.........................................................................31-Knofsimple.............................................................................. 31-Knoilfunc.................................................................................32-Knointentopt............................................................................33-Knolargepage.......................................................................... 33-Knolto......................................................................................35-Knomfunc................................................................................36-Knonf.......................................................................................36-Knons...................................................................................... 36-Knoocl..................................................................................... 36-Komitfp................................................................................... 37-Knoomitfp............................................................................... 37-Knoopenmp...................................................................... 37,344-Knooptmsg.............................................................................. 38-Knoparallel.............................................................................. 38-Knopreex................................................................................. 39-Knoprefetch.............................................................................39-Knoreduction...........................................................................42-Knoregion_extension.............................................................. 42-Knoshortloop...........................................................................42-Knosimd.................................................................................. 43-Knostatic_fjlib.........................................................................43-Knostriping..............................................................................44-Knoswp....................................................................................44-Knotemparraystack..................................................................44-Knothreadsafe...................................................................44,344-Knounroll................................................................................ 44-KNOXFILL.............................................................................45-Kns.......................................................................................... 36-Kocl......................................................................................... 36-Kopenmp.......................................................................... 37,344-Kopenmp_assume_norecurrence............................................ 37-Kopenmp_noassume_norecurrence........................................ 37-Kopenmp_noordered_reduction.......................................38,344-Kopenmp_notls................................................................ 38,344-Kopenmp_ordered_reduction...........................................37,344-Kopenmp_tls.................................................................... 38,344-Koptions.................................................................................. 38-Koptmsg........................................................................... 38,235-Kparallel........................................................................... 38,311-Kparallel_fp_precision............................................................ 38-Kparallel_iteration...................................................................39-Kparallel_nofp_precision........................................................ 39-Kparallel_strong...................................................................... 39-Kpic......................................................................................... 39-KPIC........................................................................................39-Kpreex..................................................................................... 39-Kprefetch_cache_level............................................................ 39-Kefetch_conditional................................................................ 40-Kprefetch_indirect...................................................................40-Kprefetch_infer....................................................................... 40-Kprefetch_iteration..................................................................40-Kprefetch_iteration_L2........................................................... 40-Kprefetch_line.........................................................................41-Kprefetch_line_L2.................................................................. 41-Kprefetch_noconditional.........................................................40-Kprefetch_noindirect...............................................................40

-Kprefetch_noinfer................................................................... 40-Kprefetch_nosequential...........................................................41-Kprefetch_nostride.................................................................. 41-Kprefetch_nostrong.................................................................41-K prefetch_sequential..............................................................41-Kprefetch_stride...................................................................... 41-Kprefetch_strong.....................................................................41-Kprefetch_strong_L2.............................................................. 42-Kreduction...............................................................................42-Kregion_extension.................................................................. 42-Kshortloop...............................................................................42-Ksimd...................................................................................... 43-Ksimd=1..................................................................................43-Ksimd=2..................................................................................43-Ksimd=auto............................................................................. 43-Ksimd_noseparate_stride........................................................ 43-Ksimd_separate_stride............................................................ 43-Kstatic_fjlib.............................................................................43-Kstriping..................................................................................44-Kswp........................................................................................44-Ktemparraystack......................................................................44-Kthreadsafe.......................................................................44,344-Kunroll.................................................................................... 44-Kuxsimd.................................................................................. 45-Knouxsimd.............................................................................. 45-Kvisimpact.............................................................................. 45-Kvppocl................................................................................... 45-KXFILL...................................................................................45

[L]-l...........................................................................................15,61-L.............................................................................................. 45-Lb............................................................................................ 63-le.............................................................................................. 62L edit descriptor......................................................................165-li...............................................................................................62-Li............................................................................................. 63line-feed character.................................................................. 116linker...........................................................................................4list-directed input/output statement........................................ 146little endian data......................................................................161local array................................................................................. 16logical IF statement................................................................ 106logical type data........................................................................89-Lr............................................................................................. 63-ls.............................................................................................. 62-Lu............................................................................................ 64-lw.............................................................................................62

[M]-M................................................................................. 45,64,281-m..............................................................................................62man command.............................................................................4-ml.............................................................................................15-mlcmain...................................................................................15-mldefault..................................................................................15module.................................................................................... 281

- 396 -

[N]-N.............................................................................................. 45-n...............................................................................................62-Nallextput................................................................................ 46-Nalloc_assign.......................................................................... 46namelist input/output statement..............................................139NAMELIST statement............................................................139-Ncancel_overtime_compilation.............................................. 46-Ncheck_cache_arraysize......................................................... 46-Ncheck_global.........................................................................46-Ncheck_intrfunc...................................................................... 46-Ncoarray.................................................................................. 47-Ncompdisp.............................................................................. 47-Ncopyarg................................................................................. 47-Nf90move.........................................................................47,104-Nfreealloc................................................................................ 47-Nhook_func.............................................................................48-Nhook_time.............................................................................48-Nline........................................................................................ 48-Nlst=p......................................................................................86-Nlst=t.......................................................................................86-Nmallocfree.............................................................................49-Nmaxserious............................................................................49-Nnoallextput............................................................................ 46-Nnoalloc_assign...................................................................... 46-Nnocancel_overtime_compilation.......................................... 46-Nnocoarray.............................................................................. 47-Nnocompdisp.......................................................................... 47-Nnocopyarg............................................................................. 47-Nnof90move............................................................................47-Nnofreealloc............................................................................ 47-Nnohook_func.........................................................................48-Nnohook_time.........................................................................48-Nnoline.................................................................................... 48-Nnomallocfree.........................................................................49-Nnoobsfun............................................................................... 49-Nnoprivatealloc................................................................ 49,345-Nnorecursive........................................................................... 50-Nnosetvalue.............................................................................53-Nnouse_rodata.........................................................................53-Nobsfun................................................................................... 49nonadvancing input/output statement.....................................150-Nprivatealloc......................................................................... 345-Nquickdbg[=dbg_arg]............................................................. 49-NRnotrap................................................................................. 54-NRtrap..................................................................................... 53-Nrt_notune...............................................................................51-Nrt_tune...................................................................................51-Nrt_tune_func..........................................................................51-Nrt_tune_loop..........................................................................51-Nsave.......................................................................................51-Nsetvalue.................................................................................51-Nuse_rodata.............................................................................53

[O]-o...............................................................................................15-O.............................................................................................. 54

-O0............................................................................................ 54-O1............................................................................................ 54-O2............................................................................................ 54-O3............................................................................................ 54OCL........................................................................................ 236OMP_DYNAMIC........................................................... 347,352OMP_MAX_ACTIVE_LEVELS....................................347,352OMP_NESTED............................................................... 347,352OMP_NUM_THREADS.................................................347,352OMP_PROC_BIND................................................. 345,347,352OMP_SCHEDULE..........................................................347,352OMP_STACKSIZE.................................................. 345,347,352OMP_THREAD_LIMIT................................................. 347,352OMP_WAIT_POLICY.............................................346,347,352online manual............................................................................. 4OpenMP..................................................................................343OPEN statement..................................................................... 121OPEN statement specifier.......................................................121option.......................................................................................... 4option-list....................................................................................4

[P]-P...............................................................................................55-p...............................................................................................62PAUSE statement................................................................... 107precision conversion................................................................. 93precision improving..................................................................93precision lowering and analysis of error...................................94preprocessor................................................................................2printing control command...................................................... 362PUBLIC attribute....................................................................283

[Q]-Q......................................................................................... 55,64-q...............................................................................................62-Qa............................................................................................ 55-Qd............................................................................................ 55-Qi.............................................................................................55-Qm...........................................................................................55-Qofile.......................................................................................56-Qp....................................................................................... 56,86-Qt........................................................................................56,86-Qx............................................................................................ 56

[R]-r................................................................................................62-Re............................................................................................ 64REAL and CMPLX used as generic names............................110real type data.............................................................................89RECL= Specifier.................................................................... 123relational expressions............................................................. 103relational operation................................................................. 103removing a common expression............................................. 229REWIND Statement............................................................... 154REWIND statement................................................................158rounding error........................................................................... 18-Rp............................................................................................ 64runtime option...........................................................................60

- 397 -

-Ry............................................................................................ 64

[S]-S...............................................................................................56SAVE attribute for initialized variable...................................102SELECT TYPE construct.......................................................359sequential access.....................................................................120service routine.........................................................................111-shared...................................................................................... 15size of an array element.......................................................... 101-SSL2........................................................................................ 56-SSL2BLAMP.......................................................................... 56STAT= specifier..................................................................... 108statement function reference...................................................110STATUS= Specifier............................................................... 123STOP statement...................................................................... 107substring......................................................................25,205,359

[T]-t................................................................................................63-T.............................................................................................. 65THREAD_STACK_SIZE...................................................... 345TMPDIR................................................................................... 58trace back map.......................................................................... 87

[U]-U.............................................................................................. 56unformatted sequential input/output statement.......................137USE statement........................................................................ 281using the optimization control line......................................... 236

[V]-v...............................................................................................15-V.............................................................................................. 56-v03d.........................................................................................15-v03e......................................................................................... 15-v03o.........................................................................................16-v03s......................................................................................... 16version and release information................................................56

[W]-W............................................................................................. 56-Wl............................................................................................ 60

[X]-x...................................................................................16,63,232-X.............................................................................................. 57-x-..............................................................................................16-X03.......................................................................................... 57-X6............................................................................................ 57-X7............................................................................................ 57-X9............................................................................................ 57-Xd7.......................................................................................... 57-xdat_szK..................................................................................16-xdir=dir_name......................................................................... 16-xproc_name............................................................................. 16-xstmt_no.................................................................................. 16

- 398 -